xref: /titanic_41/usr/src/uts/common/io/lvm/raid/raid.c (revision 35a5a3587fd94b666239c157d3722745250ccbd7)
1 /*
2  * CDDL HEADER START
3  *
4  * The contents of this file are subject to the terms of the
5  * Common Development and Distribution License (the "License").
6  * You may not use this file except in compliance with the License.
7  *
8  * You can obtain a copy of the license at usr/src/OPENSOLARIS.LICENSE
9  * or http://www.opensolaris.org/os/licensing.
10  * See the License for the specific language governing permissions
11  * and limitations under the License.
12  *
13  * When distributing Covered Code, include this CDDL HEADER in each
14  * file and include the License file at usr/src/OPENSOLARIS.LICENSE.
15  * If applicable, add the following below this CDDL HEADER, with the
16  * fields enclosed by brackets "[]" replaced with your own identifying
17  * information: Portions Copyright [yyyy] [name of copyright owner]
18  *
19  * CDDL HEADER END
20  */
21 
22 /*
23  * Copyright 2009 Sun Microsystems, Inc.  All rights reserved.
24  * Use is subject to license terms.
25  */
26 
27 /*
28  * NAME:	raid.c
29  *
30  * DESCRIPTION: Main RAID driver source file containing open, close and I/O
31  *		operations.
32  *
33  * ROUTINES PROVIDED FOR EXTERNAL USE:
34  *  raid_open()			- open the RAID metadevice for access.
35  *  raid_internal_open()	- internal open routine of RAID metdevice.
36  *  md_raid_strategy()		- perform normal I/O operations,
37  *				    such as read and write.
38  *  raid_close()		- close the RAID metadevice.
39  *  raid_internal_close()	- internal close routine of RAID metadevice.
40  *  raid_snarf()		- initialize and clean up MDD records.
41  *  raid_halt()			- reset the RAID metadevice
42  *  raid_line()			- return the line # of this segment
43  *  raid_dcolumn()		- return the data column # of this segment
44  *  raid_pcolumn()		- return the parity column # of this segment
45  */
46 
47 #include <sys/param.h>
48 #include <sys/systm.h>
49 #include <sys/conf.h>
50 #include <sys/file.h>
51 #include <sys/user.h>
52 #include <sys/uio.h>
53 #include <sys/t_lock.h>
54 #include <sys/buf.h>
55 #include <sys/dkio.h>
56 #include <sys/vtoc.h>
57 #include <sys/kmem.h>
58 #include <vm/page.h>
59 #include <sys/cmn_err.h>
60 #include <sys/sysmacros.h>
61 #include <sys/types.h>
62 #include <sys/mkdev.h>
63 #include <sys/stat.h>
64 #include <sys/open.h>
65 #include <sys/modctl.h>
66 #include <sys/ddi.h>
67 #include <sys/sunddi.h>
68 #include <sys/debug.h>
69 #include <sys/lvm/md_raid.h>
70 #include <sys/lvm/mdvar.h>
71 #include <sys/lvm/md_convert.h>
72 
73 #include <sys/sysevent/eventdefs.h>
74 #include <sys/sysevent/svm.h>
75 
76 md_ops_t		raid_md_ops;
77 #ifndef lint
78 char			_depends_on[] = "drv/md";
79 md_ops_t		*md_interface_ops = &raid_md_ops;
80 #endif	/* lint */
81 
82 extern unit_t		md_nunits;
83 extern unit_t		md_nsets;
84 extern md_set_t		md_set[];
85 extern int		md_status;
86 extern major_t		md_major;
87 extern mdq_anchor_t	md_done_daemon;
88 extern mdq_anchor_t	md_mstr_daemon;
89 extern int		md_sleep_for_test;
90 extern clock_t		md_hz;
91 
92 extern md_event_queue_t	*md_event_queue;
93 
94 
95 int pchunks		= 16;
96 int phigh		= 1024;
97 int plow		= 128;
98 int cchunks		= 64;
99 int chigh		= 1024;
100 int clow		= 512;
101 int bchunks		= 32;
102 int bhigh		= 256;
103 int blow		= 128;
104 
105 int raid_total_io		= 0;
106 int raid_reads			= 0;
107 int raid_writes			= 0;
108 int raid_no_bpmaps		= 0;
109 int raid_512			= 0;
110 int raid_1024			= 0;
111 int raid_1024_8192		= 0;
112 int raid_8192			= 0;
113 int raid_8192_bigger		= 0;
114 int raid_line_lock_wait	= 0;
115 
116 int data_buffer_waits		= 0;
117 int parity_buffer_waits	= 0;
118 
119 /* writer line locks */
120 int raid_writer_locks		= 0; /* total writer locks */
121 int raid_write_waits		= 0; /* total writer locks that waited */
122 int raid_full_line_writes	= 0; /* total full line writes */
123 int raid_write_queue_length	= 0; /* wait queue length */
124 int raid_max_write_q_length	= 0; /* maximum queue length */
125 int raid_write_locks_active	= 0; /* writer locks at any time */
126 int raid_max_write_locks	= 0; /* maximum writer locks active */
127 
128 /* read line locks */
129 int raid_reader_locks		= 0; /* total reader locks held */
130 int raid_reader_locks_active	= 0; /* reader locks held */
131 int raid_max_reader_locks	= 0; /* maximum reader locks held in run */
132 int raid_read_overlaps		= 0; /* number of times 2 reads hit same line */
133 int raid_read_waits		= 0; /* times a reader waited on writer */
134 
135 /* prewrite stats */
136 int raid_prewrite_waits		= 0; /* number of waits for a pw slot */
137 int raid_pw			= 0; /* number of pw slots in use */
138 int raid_prewrite_max		= 0; /* maximum number of pw slots in use */
139 int raid_pw_invalidates		= 0;
140 
141 static clock_t md_wr_wait	= 0;
142 
143 int nv_available	= 0; /* presence of nv-ram support in device */
144 int nv_prewrite		= 1; /* mark prewrites with nv_available */
145 int nv_parity		= 1; /* mark parity with nv_available */
146 
147 kmem_cache_t	*raid_parent_cache = NULL;
148 kmem_cache_t	*raid_child_cache = NULL;
149 kmem_cache_t	*raid_cbuf_cache = NULL;
150 
151 int			raid_internal_open(minor_t mnum, int flag, int otyp,
152 			    int md_oflags);
153 
154 static void		freebuffers(md_raidcs_t *cs);
155 static int		raid_read(mr_unit_t *un, md_raidcs_t *cs);
156 static void		raid_read_io(mr_unit_t *un, md_raidcs_t *cs);
157 static int		raid_write(mr_unit_t *un, md_raidcs_t *cs);
158 static void		raid_write_io(mr_unit_t *un, md_raidcs_t *cs);
159 static void		raid_stage(md_raidcs_t *cs);
160 static void		raid_enqueue(md_raidcs_t *cs);
161 static diskaddr_t	raid_line(diskaddr_t segment, mr_unit_t *un);
162 uint_t			raid_dcolumn(diskaddr_t segment, mr_unit_t *un);
163 static void		getpbuffer(md_raidcs_t *cs);
164 static void		getdbuffer(md_raidcs_t *cs);
165 static void		raid_done(buf_t *bp);
166 static void		raid_io_startup(mr_unit_t *un);
167 
168 static rus_state_t
169 raid_col2unit(rcs_state_t state, rus_state_t unitstate)
170 {
171 	switch (state) {
172 	case RCS_INIT:
173 		return (RUS_INIT);
174 	case RCS_OKAY:
175 		return (RUS_OKAY);
176 	case RCS_RESYNC:
177 		if (unitstate & RUS_LAST_ERRED)
178 			return (RUS_LAST_ERRED);
179 		else
180 			return (RUS_ERRED);
181 	case RCS_ERRED:
182 		return (RUS_ERRED);
183 	case RCS_LAST_ERRED:
184 		return (RUS_ERRED);
185 	default:
186 		break;
187 	}
188 	panic("raid_col2unit");
189 	/*NOTREACHED*/
190 }
191 
192 void
193 raid_set_state(mr_unit_t *un, int col, rcs_state_t newstate, int force)
194 {
195 
196 	rus_state_t	unitstate, origstate;
197 	rcs_state_t	colstate;
198 	rcs_state_t	orig_colstate;
199 	int		errcnt = 0, okaycnt = 0, resynccnt = 0;
200 	int		i;
201 	char		*devname;
202 
203 	ASSERT(un);
204 	ASSERT(col < un->un_totalcolumncnt);
205 	ASSERT(newstate &
206 	    (RCS_INIT | RCS_INIT_ERRED | RCS_OKAY | RCS_RESYNC | RCS_ERRED |
207 	    RCS_LAST_ERRED | RCS_REGEN));
208 	ASSERT((newstate &
209 	    ~(RCS_INIT | RCS_INIT_ERRED | RCS_OKAY | RCS_RESYNC | RCS_ERRED |
210 	    RCS_LAST_ERRED | RCS_REGEN))
211 	    == 0);
212 
213 	ASSERT(MDI_UNIT(MD_SID(un)) ? UNIT_WRITER_HELD(un) : 1);
214 
215 	unitstate = un->un_state;
216 	origstate = unitstate;
217 
218 	if (force) {
219 		un->un_column[col].un_devstate = newstate;
220 		un->un_state = raid_col2unit(newstate, unitstate);
221 		uniqtime32(&un->un_column[col].un_devtimestamp);
222 		uniqtime32(&un->un_timestamp);
223 		return;
224 	}
225 
226 	ASSERT(un->un_state &
227 	    (RUS_INIT | RUS_OKAY | RUS_ERRED | RUS_DOI | RUS_LAST_ERRED |
228 	    RUS_REGEN));
229 	ASSERT((un->un_state & ~(RUS_INIT |
230 	    RUS_OKAY | RUS_ERRED | RUS_DOI | RUS_LAST_ERRED | RUS_REGEN)) == 0);
231 
232 	if (un->un_column[col].un_devstate == newstate)
233 		return;
234 
235 	if (newstate == RCS_REGEN) {
236 		if (raid_state_cnt(un, RCS_OKAY) != un->un_totalcolumncnt)
237 			return;
238 		un->un_state = RUS_REGEN;
239 		return;
240 	}
241 
242 	orig_colstate = un->un_column[col].un_devstate;
243 
244 	/*
245 	 * if there is another column in the error state then this
246 	 * column should go to the last errored state
247 	 */
248 	for (i = 0; i < un->un_totalcolumncnt; i++) {
249 		if (i == col)
250 			colstate = newstate;
251 		else
252 			colstate = un->un_column[i].un_devstate;
253 		if (colstate & (RCS_ERRED | RCS_LAST_ERRED | RCS_INIT_ERRED))
254 			errcnt++;
255 		if (colstate & RCS_OKAY)
256 			okaycnt++;
257 		if (colstate & RCS_RESYNC)
258 			resynccnt++;
259 	}
260 	ASSERT(resynccnt < 2);
261 
262 	if (okaycnt == un->un_totalcolumncnt)
263 		unitstate = RUS_OKAY;
264 	else if (errcnt > 1) {
265 		unitstate = RUS_LAST_ERRED;
266 		if (newstate & RCS_ERRED)
267 			newstate = RCS_LAST_ERRED;
268 	} else if (errcnt == 1)
269 		if (!(unitstate & RUS_LAST_ERRED))
270 			unitstate = RUS_ERRED;
271 
272 	if (un->un_state == RUS_DOI)
273 		unitstate = RUS_DOI;
274 
275 	un->un_column[col].un_devstate = newstate;
276 	uniqtime32(&un->un_column[col].un_devtimestamp);
277 	/*
278 	 * if there are last errored column being brought back online
279 	 * by open or snarf, then be sure to clear the RUS_LAST_ERRED
280 	 * bit to allow writes.  If there is a real error then the
281 	 * column will go back into last erred.
282 	 */
283 	if ((raid_state_cnt(un, RCS_LAST_ERRED) == 0) &&
284 	    (raid_state_cnt(un, RCS_ERRED) == 1))
285 		unitstate = RUS_ERRED;
286 
287 	un->un_state = unitstate;
288 	uniqtime32(&un->un_timestamp);
289 
290 	if ((! (origstate & (RUS_ERRED|RUS_LAST_ERRED|RUS_DOI))) &&
291 	    (unitstate & (RUS_ERRED|RUS_LAST_ERRED|RUS_DOI))) {
292 		devname = md_devname(MD_UN2SET(un),
293 		    un->un_column[col].un_dev, NULL, 0);
294 
295 		cmn_err(CE_WARN, "md: %s: %s needs maintenance",
296 		    md_shortname(MD_SID(un)), devname);
297 
298 		if (unitstate & RUS_LAST_ERRED) {
299 			cmn_err(CE_WARN, "md: %s: %s last erred",
300 			    md_shortname(MD_SID(un)), devname);
301 
302 		} else if (un->un_column[col].un_devflags &
303 		    MD_RAID_DEV_ISOPEN) {
304 			/*
305 			 * Close the broken device and clear the open flag on
306 			 * it.  We have to check that the device is open,
307 			 * otherwise the first open on it has resulted in the
308 			 * error that is being processed and the actual un_dev
309 			 * will be NODEV64.
310 			 */
311 			md_layered_close(un->un_column[col].un_dev,
312 			    MD_OFLG_NULL);
313 			un->un_column[col].un_devflags &= ~MD_RAID_DEV_ISOPEN;
314 		}
315 	} else if (orig_colstate == RCS_LAST_ERRED && newstate == RCS_ERRED &&
316 	    un->un_column[col].un_devflags & MD_RAID_DEV_ISOPEN) {
317 		/*
318 		 * Similar to logic above except no log messages since we
319 		 * are just transitioning from Last Erred to Erred.
320 		 */
321 		md_layered_close(un->un_column[col].un_dev, MD_OFLG_NULL);
322 		un->un_column[col].un_devflags &= ~MD_RAID_DEV_ISOPEN;
323 	}
324 
325 	/*
326 	 * If a resync has completed, see if there is a Last Erred
327 	 * component that we can change to the Erred state.
328 	 */
329 	if ((orig_colstate == RCS_RESYNC) && (newstate == RCS_OKAY)) {
330 		for (i = 0; i < un->un_totalcolumncnt; i++) {
331 			if (i != col &&
332 			    (un->un_column[i].un_devstate & RCS_LAST_ERRED)) {
333 				raid_set_state(un, i, RCS_ERRED, 0);
334 				break;
335 			}
336 		}
337 	}
338 }
339 
340 /*
341  * NAME:	erred_check_line
342  *
343  * DESCRIPTION: Return the type of write to perform on an erred column based
344  *		upon any resync activity.
345  *
346  *		if a column is being resynced and the write is above the
347  *		resync point may have to write to the target being resynced.
348  *
349  *		Column state may make it impossible to do the write
350  *		in which case RCL_EIO or RCL_ENXIO is returned.
351  *
352  *		If a column cannot be written directly, RCL_ERRED is
353  *		returned and processing should proceed accordingly.
354  *
355  * PARAMETERS:	minor_t		 mnum - minor number identity of metadevice
356  *		md_raidcs_t	 *cs - child save structure
357  *		mr_column_t	 *dcolumn - pointer to data column structure
358  *		mr_column_t	 *pcolumn - pointer to parity column structure
359  *
360  * RETURNS:	RCL_OKAY, RCL_ERRED
361  *
362  * LOCKS:	Expects Line Writer Lock and Unit Resource Lock to be held
363  *		across call.
364  */
365 
366 static int
367 erred_check_line(mr_unit_t *un, md_raidcs_t *cs, mr_column_t *column)
368 {
369 
370 	ASSERT(un != NULL);
371 	ASSERT(cs->cs_flags & MD_RCS_LLOCKD);
372 
373 	if (column->un_devstate & RCS_OKAY)
374 		return (RCL_OKAY);
375 
376 	if (column->un_devstate & RCS_ERRED)
377 		return (RCL_ERRED);  /* do not read from errored disk */
378 
379 	/*
380 	 * for the last errored case their are two considerations.
381 	 * When the last errored column is the only errored column then
382 	 * do treat it like a maintenance column, not doing I/O from
383 	 * it.   When it there are other failures then just attempt
384 	 * to use it.
385 	 */
386 	if (column->un_devstate & RCS_LAST_ERRED)
387 		return (RCL_ERRED);
388 
389 	ASSERT(column->un_devstate & RCS_RESYNC);
390 
391 	/*
392 	 * When a resync from a hotspare is being done (copy resync)
393 	 * then always treat it as an OKAY column, since no regen
394 	 * is required.
395 	 */
396 	if (column->un_devflags & MD_RAID_COPY_RESYNC) {
397 		return (RCL_OKAY);
398 	}
399 
400 	mutex_enter(&un->un_mx);
401 	if (cs->cs_line < un->un_resync_line_index) {
402 		mutex_exit(&un->un_mx);
403 		return (RCL_OKAY);
404 	}
405 	mutex_exit(&un->un_mx);
406 	return (RCL_ERRED);
407 
408 }
409 
410 /*
411  * NAMES:	raid_state_cnt
412  *
413  * DESCRIPTION: counts number of column in a specific state
414  *
415  * PARAMETERS:	md_raid_t *un
416  *		rcs_state state
417  */
418 int
419 raid_state_cnt(mr_unit_t *un, rcs_state_t state)
420 {
421 	int	i, retval = 0;
422 
423 	for (i = 0; i < un->un_totalcolumncnt; i++)
424 		if (un->un_column[i].un_devstate & state)
425 			retval++;
426 	return (retval);
427 }
428 
429 /*
430  * NAMES:	raid_io_overlaps
431  *
432  * DESCRIPTION: checkst for overlap of 2 child save structures
433  *
434  * PARAMETERS:	md_raidcs_t cs1
435  *		md_raidcs_t cs2
436  *
437  * RETURNS:	0 - no overlap
438  *		1 - overlap
439  */
440 int
441 raid_io_overlaps(md_raidcs_t *cs1, md_raidcs_t *cs2)
442 {
443 	if (cs1->cs_blkno > cs2->cs_lastblk)
444 		return (0);
445 	if (cs1->cs_lastblk < cs2->cs_blkno)
446 		return (0);
447 	return (1);
448 }
449 
450 /*
451  * NAMES:	raid_parent_constructor
452  * DESCRIPTION: parent structure constructor routine
453  * PARAMETERS:
454  */
455 /*ARGSUSED1*/
456 static int
457 raid_parent_constructor(void *p, void *d1, int d2)
458 {
459 	mutex_init(&((md_raidps_t *)p)->ps_mx,
460 	    NULL, MUTEX_DEFAULT, NULL);
461 	mutex_init(&((md_raidps_t *)p)->ps_mapin_mx,
462 	    NULL, MUTEX_DEFAULT, NULL);
463 	return (0);
464 }
465 
466 void
467 raid_parent_init(md_raidps_t *ps)
468 {
469 	bzero(ps, offsetof(md_raidps_t, ps_mx));
470 	((md_raidps_t *)ps)->ps_flags = MD_RPS_INUSE;
471 	((md_raidps_t *)ps)->ps_magic = RAID_PSMAGIC;
472 }
473 
474 /*ARGSUSED1*/
475 static void
476 raid_parent_destructor(void *p, void *d)
477 {
478 	mutex_destroy(&((md_raidps_t *)p)->ps_mx);
479 	mutex_destroy(&((md_raidps_t *)p)->ps_mapin_mx);
480 }
481 
482 /*
483  * NAMES:	raid_child_constructor
484  * DESCRIPTION: child structure constructor routine
485  * PARAMETERS:
486  */
487 /*ARGSUSED1*/
488 static int
489 raid_child_constructor(void *p, void *d1, int d2)
490 {
491 	md_raidcs_t	*cs = (md_raidcs_t *)p;
492 	mutex_init(&cs->cs_mx, NULL, MUTEX_DEFAULT, NULL);
493 	bioinit(&cs->cs_dbuf);
494 	bioinit(&cs->cs_pbuf);
495 	bioinit(&cs->cs_hbuf);
496 	return (0);
497 }
498 
499 void
500 raid_child_init(md_raidcs_t *cs)
501 {
502 	bzero(cs, offsetof(md_raidcs_t, cs_mx));
503 
504 	md_bioreset(&cs->cs_dbuf);
505 	md_bioreset(&cs->cs_pbuf);
506 	md_bioreset(&cs->cs_hbuf);
507 
508 	((md_raidcs_t *)cs)->cs_dbuf.b_chain =
509 	    ((md_raidcs_t *)cs)->cs_pbuf.b_chain =
510 	    ((md_raidcs_t *)cs)->cs_hbuf.b_chain =
511 	    (struct buf *)(cs);
512 
513 	cs->cs_magic = RAID_CSMAGIC;
514 	cs->cs_line = MD_DISKADDR_ERROR;
515 	cs->cs_dpwslot = -1;
516 	cs->cs_ppwslot = -1;
517 }
518 
519 /*ARGSUSED1*/
520 static void
521 raid_child_destructor(void *p, void *d)
522 {
523 	biofini(&((md_raidcs_t *)p)->cs_dbuf);
524 	biofini(&((md_raidcs_t *)p)->cs_hbuf);
525 	biofini(&((md_raidcs_t *)p)->cs_pbuf);
526 	mutex_destroy(&((md_raidcs_t *)p)->cs_mx);
527 }
528 
529 /*ARGSUSED1*/
530 static int
531 raid_cbuf_constructor(void *p, void *d1, int d2)
532 {
533 	bioinit(&((md_raidcbuf_t *)p)->cbuf_bp);
534 	return (0);
535 }
536 
537 static void
538 raid_cbuf_init(md_raidcbuf_t *cb)
539 {
540 	bzero(cb, offsetof(md_raidcbuf_t, cbuf_bp));
541 	md_bioreset(&cb->cbuf_bp);
542 	cb->cbuf_magic = RAID_BUFMAGIC;
543 	cb->cbuf_pwslot = -1;
544 	cb->cbuf_flags = CBUF_WRITE;
545 }
546 
547 /*ARGSUSED1*/
548 static void
549 raid_cbuf_destructor(void *p, void *d)
550 {
551 	biofini(&((md_raidcbuf_t *)p)->cbuf_bp);
552 }
553 
554 /*
555  * NAMES:	raid_run_queue
556  * DESCRIPTION: spawn a backend processing daemon for RAID metadevice.
557  * PARAMETERS:
558  */
559 /*ARGSUSED*/
560 static void
561 raid_run_queue(void *d)
562 {
563 	if (!(md_status & MD_GBL_DAEMONS_LIVE))
564 		md_daemon(1, &md_done_daemon);
565 }
566 
567 /*
568  * NAME:	raid_build_pwslot
569  * DESCRIPTION: builds mr_pw_reserve for the column
570  * PARAMETERS:	un is the pointer to the unit structure
571  *		colindex is the column to create the structure for
572  */
573 int
574 raid_build_pw_reservation(mr_unit_t *un, int colindex)
575 {
576 	mr_pw_reserve_t	*pw;
577 	mr_scoreboard_t	*sb;
578 	int		i;
579 
580 	pw = (mr_pw_reserve_t *) kmem_zalloc(sizeof (mr_pw_reserve_t) +
581 	    (sizeof (mr_scoreboard_t) * un->un_pwcnt), KM_SLEEP);
582 	pw->pw_magic = RAID_PWMAGIC;
583 	pw->pw_column = colindex;
584 	pw->pw_free = un->un_pwcnt;
585 	sb = &pw->pw_sb[0];
586 	for (i = 0; i < un->un_pwcnt; i++) {
587 		sb[i].sb_column = colindex;
588 		sb[i].sb_flags = SB_UNUSED;
589 		sb[i].sb_start_blk = 0;
590 		sb[i].sb_last_blk = 0;
591 		sb[i].sb_cs = NULL;
592 	}
593 	un->un_column_ic[colindex].un_pw_reserve = pw;
594 	return (0);
595 }
596 /*
597  * NAME:	raid_free_pw_reservation
598  * DESCRIPTION: RAID metadevice pre-write slot structure destroy routine
599  * PARAMETERS:	mr_unit_t *un - pointer to a unit structure
600  *		int colindex  - index of the column whose pre-write slot struct
601  *			is to be destroyed.
602  */
603 void
604 raid_free_pw_reservation(mr_unit_t *un, int colindex)
605 {
606 	mr_pw_reserve_t	*pw = un->un_column_ic[colindex].un_pw_reserve;
607 
608 	kmem_free(pw, sizeof (mr_pw_reserve_t) +
609 	    (sizeof (mr_scoreboard_t) * un->un_pwcnt));
610 }
611 
612 /*
613  * NAME:	raid_cancel_pwslot
614  * DESCRIPTION: RAID metadevice write routine
615  * PARAMETERS:	md_raidcs_t *cs - pointer to a child structure
616  */
617 static void
618 raid_cancel_pwslot(md_raidcs_t *cs)
619 {
620 	mr_unit_t		*un = cs->cs_un;
621 	mr_pw_reserve_t		*pw;
622 	mr_scoreboard_t		*sb;
623 	mr_column_ic_t		*col;
624 	md_raidcbuf_t		*cbuf;
625 	int			broadcast = 0;
626 
627 	if (cs->cs_ps->ps_flags & MD_RPS_READ)
628 		return;
629 	if (cs->cs_dpwslot != -1) {
630 		col = &un->un_column_ic[cs->cs_dcolumn];
631 		pw = col->un_pw_reserve;
632 		sb = &pw->pw_sb[cs->cs_dpwslot];
633 		sb->sb_flags = SB_AVAIL;
634 		if ((pw->pw_free++ == 0) || (un->un_rflags & MD_RFLAG_NEEDPW))
635 			broadcast++;
636 		sb->sb_cs = NULL;
637 	}
638 
639 	if (cs->cs_ppwslot != -1) {
640 		col = &un->un_column_ic[cs->cs_pcolumn];
641 		pw = col->un_pw_reserve;
642 		sb = &pw->pw_sb[cs->cs_ppwslot];
643 		sb->sb_flags = SB_AVAIL;
644 		if ((pw->pw_free++ == 0) || (un->un_rflags & MD_RFLAG_NEEDPW))
645 			broadcast++;
646 		sb->sb_cs = NULL;
647 	}
648 
649 	for (cbuf = cs->cs_buflist; cbuf; cbuf = cbuf->cbuf_next) {
650 		if (cbuf->cbuf_pwslot == -1)
651 			continue;
652 		col = &un->un_column_ic[cbuf->cbuf_column];
653 		pw = col->un_pw_reserve;
654 		sb = &pw->pw_sb[cbuf->cbuf_pwslot];
655 		sb->sb_flags = SB_AVAIL;
656 		if ((pw->pw_free++ == 0) || (un->un_rflags & MD_RFLAG_NEEDPW))
657 			broadcast++;
658 		sb->sb_cs = NULL;
659 	}
660 	if (broadcast) {
661 		cv_broadcast(&un->un_cv);
662 		return;
663 	}
664 	mutex_enter(&un->un_mx);
665 	if (un->un_rflags & MD_RFLAG_NEEDPW)
666 		cv_broadcast(&un->un_cv);
667 	mutex_exit(&un->un_mx);
668 }
669 
670 static void
671 raid_free_pwinvalidate(md_raidcs_t *cs)
672 {
673 	md_raidcbuf_t		*cbuf;
674 	md_raidcbuf_t		*cbuf_to_free;
675 	mr_unit_t		*un = cs->cs_un;
676 	mdi_unit_t		*ui = MDI_UNIT(MD_SID(un));
677 	mr_pw_reserve_t		*pw;
678 	mr_scoreboard_t		*sb;
679 	int			broadcast = 0;
680 
681 	cbuf = cs->cs_pw_inval_list;
682 	ASSERT(cbuf);
683 	mutex_enter(&un->un_linlck_mx);
684 	while (cbuf) {
685 		pw = un->un_column_ic[cbuf->cbuf_column].un_pw_reserve;
686 		sb = &pw->pw_sb[0];
687 		ASSERT(sb[cbuf->cbuf_pwslot].sb_flags & SB_INVAL_PEND);
688 		sb[cbuf->cbuf_pwslot].sb_flags = SB_UNUSED;
689 		sb[cbuf->cbuf_pwslot].sb_cs = NULL;
690 		if ((pw->pw_free++ == 0) || (un->un_rflags & MD_RFLAG_NEEDPW))
691 			broadcast++;
692 		cbuf_to_free = cbuf;
693 		cbuf = cbuf->cbuf_next;
694 		kmem_free(cbuf_to_free->cbuf_buffer, dbtob(un->un_iosize));
695 		kmem_cache_free(raid_cbuf_cache, cbuf_to_free);
696 	}
697 	cs->cs_pw_inval_list = (md_raidcbuf_t *)NULL;
698 	/*
699 	 * now that there is a free prewrite slot, check to see if there
700 	 * are any io operations waiting first wake up the raid_io_startup
701 	 * then signal the the processes waiting in raid_write.
702 	 */
703 	if (ui->ui_io_lock->io_list_front)
704 		raid_io_startup(un);
705 	mutex_exit(&un->un_linlck_mx);
706 	if (broadcast) {
707 		cv_broadcast(&un->un_cv);
708 		return;
709 	}
710 	mutex_enter(&un->un_mx);
711 	if (un->un_rflags & MD_RFLAG_NEEDPW)
712 		cv_broadcast(&un->un_cv);
713 	mutex_exit(&un->un_mx);
714 }
715 
716 
717 static int
718 raid_get_pwslot(md_raidcs_t *cs, int column)
719 {
720 	mr_scoreboard_t	*sb;
721 	mr_pw_reserve_t	*pw;
722 	mr_unit_t	*un = cs->cs_un;
723 	diskaddr_t	start_blk = cs->cs_blkno;
724 	diskaddr_t	last_blk = cs->cs_lastblk;
725 	int		i;
726 	int		pwcnt = un->un_pwcnt;
727 	int		avail = -1;
728 	int		use = -1;
729 	int		flags;
730 
731 
732 	/* start with the data column */
733 	pw = cs->cs_un->un_column_ic[column].un_pw_reserve;
734 	sb = &pw->pw_sb[0];
735 	ASSERT(pw->pw_free > 0);
736 	for (i = 0; i < pwcnt; i++) {
737 		flags = sb[i].sb_flags;
738 		if (flags & SB_INVAL_PEND)
739 			continue;
740 
741 		if ((avail == -1) && (flags & (SB_AVAIL | SB_UNUSED)))
742 			avail = i;
743 
744 		if ((start_blk > sb[i].sb_last_blk) ||
745 		    (last_blk < sb[i].sb_start_blk))
746 			continue;
747 
748 		/* OVERLAP */
749 		ASSERT(! (sb[i].sb_flags & SB_INUSE));
750 
751 		/*
752 		 * raid_invalidate_pwslot attempts to zero out prewrite entry
753 		 * in parallel with other disk reads/writes related to current
754 		 * transaction. however cs_frags accounting for this case is
755 		 * broken because raid_write_io resets cs_frags i.e. ignoring
756 		 * that it could have been been set to > 0 value by
757 		 * raid_invalidate_pwslot. While this can be fixed an
758 		 * additional problem is that we don't seem to handle
759 		 * correctly the case of getting a disk error for prewrite
760 		 * entry invalidation.
761 		 * It does not look like we really need
762 		 * to invalidate prewrite slots because raid_replay sorts
763 		 * prewrite id's in ascending order and during recovery the
764 		 * latest prewrite entry for the same block will be replay
765 		 * last. That's why i ifdef'd out the call to
766 		 * raid_invalidate_pwslot. --aguzovsk@east
767 		 */
768 
769 		if (use == -1) {
770 			use = i;
771 		}
772 	}
773 
774 	ASSERT(avail != -1);
775 	pw->pw_free--;
776 	if (use == -1)
777 		use = avail;
778 
779 	ASSERT(! (sb[use].sb_flags & SB_INUSE));
780 	sb[use].sb_flags = SB_INUSE;
781 	sb[use].sb_cs = cs;
782 	sb[use].sb_start_blk = start_blk;
783 	sb[use].sb_last_blk = last_blk;
784 	ASSERT((use >= 0) && (use < un->un_pwcnt));
785 	return (use);
786 }
787 
788 static int
789 raid_check_pw(md_raidcs_t *cs)
790 {
791 
792 	mr_unit_t	*un = cs->cs_un;
793 	int		i;
794 
795 	ASSERT(! (cs->cs_flags & MD_RCS_HAVE_PW_SLOTS));
796 	/*
797 	 * check to be sure there is a prewrite slot available
798 	 * if not just return.
799 	 */
800 	if (cs->cs_flags & MD_RCS_LINE) {
801 		for (i = 0; i < un->un_totalcolumncnt; i++)
802 			if (un->un_column_ic[i].un_pw_reserve->pw_free <= 0)
803 				return (1);
804 		return (0);
805 	}
806 
807 	if (un->un_column_ic[cs->cs_dcolumn].un_pw_reserve->pw_free <= 0)
808 		return (1);
809 	if (un->un_column_ic[cs->cs_pcolumn].un_pw_reserve->pw_free <= 0)
810 		return (1);
811 	return (0);
812 }
813 static int
814 raid_alloc_pwslot(md_raidcs_t *cs)
815 {
816 	mr_unit_t	*un = cs->cs_un;
817 	md_raidcbuf_t	*cbuf;
818 
819 	ASSERT(! (cs->cs_flags & MD_RCS_HAVE_PW_SLOTS));
820 	if (raid_check_pw(cs))
821 		return (1);
822 
823 	mutex_enter(&un->un_mx);
824 	un->un_pwid++;
825 	cs->cs_pwid = un->un_pwid;
826 	mutex_exit(&un->un_mx);
827 
828 	cs->cs_dpwslot = raid_get_pwslot(cs, cs->cs_dcolumn);
829 	for (cbuf = cs->cs_buflist; cbuf; cbuf = cbuf->cbuf_next) {
830 		cbuf->cbuf_pwslot = raid_get_pwslot(cs, cbuf->cbuf_column);
831 	}
832 	cs->cs_ppwslot = raid_get_pwslot(cs, cs->cs_pcolumn);
833 
834 	cs->cs_flags |= MD_RCS_HAVE_PW_SLOTS;
835 
836 	return (0);
837 }
838 
839 /*
840  * NAMES:	raid_build_incore
841  * DESCRIPTION: RAID metadevice incore structure building routine
842  * PARAMETERS:	void *p - pointer to a unit structure
843  *		int snarfing - a flag to indicate snarfing is required
844  */
845 int
846 raid_build_incore(void *p, int snarfing)
847 {
848 	mr_unit_t	*un = (mr_unit_t *)p;
849 	minor_t		mnum = MD_SID(un);
850 	mddb_recid_t	hs_recid = 0;
851 	int		i;
852 	int		preserve_flags;
853 	mr_column_t	*column;
854 	int		iosize;
855 	md_dev64_t	hs, dev;
856 	int		resync_cnt = 0, error_cnt = 0;
857 
858 	hs = NODEV64;
859 	dev = NODEV64;
860 
861 	/* clear out bogus pointer incase we return(1) prior to alloc */
862 	un->mr_ic = NULL;
863 
864 	if (MD_STATUS(un) & MD_UN_BEING_RESET) {
865 		mddb_setrecprivate(un->c.un_record_id, MD_PRV_PENDCLEAN);
866 		return (1);
867 	}
868 
869 	if (MD_UNIT(mnum) != NULL)
870 		return (0);
871 
872 	if (snarfing)
873 		MD_STATUS(un) = 0;
874 
875 	un->mr_ic = (mr_unit_ic_t *)kmem_zalloc(sizeof (*un->mr_ic),
876 	    KM_SLEEP);
877 
878 	un->un_column_ic = (mr_column_ic_t *)
879 	    kmem_zalloc(sizeof (mr_column_ic_t) *
880 	    un->un_totalcolumncnt, KM_SLEEP);
881 
882 	for (i = 0; i < un->un_totalcolumncnt; i++) {
883 
884 		column	= &un->un_column[i];
885 		preserve_flags = column->un_devflags &
886 		    (MD_RAID_COPY_RESYNC | MD_RAID_REGEN_RESYNC);
887 		column->un_devflags &=
888 		    ~(MD_RAID_ALT_ISOPEN | MD_RAID_DEV_ISOPEN |
889 		    MD_RAID_WRITE_ALT);
890 		if (raid_build_pw_reservation(un, i) != 0) {
891 			/* could not build pwslot */
892 			return (1);
893 		}
894 
895 		if (snarfing) {
896 			set_t		setno = MD_MIN2SET(mnum);
897 			dev =  md_getdevnum(setno, mddb_getsidenum(setno),
898 			    column->un_orig_key, MD_NOTRUST_DEVT);
899 			/*
900 			 * Comment out instead of remove so we have history
901 			 * In the pre-SVM releases stored devt is used so
902 			 * as long as there is one snarf is always happy
903 			 * even the component is powered off.  This is not
904 			 * the case in current SVM implementation.  NODEV64
905 			 * can be returned and in this case since we resolve
906 			 * the devt at 'open' time (first use of metadevice)
907 			 * we will allow snarf continue.
908 			 *
909 			 * if (dev == NODEV64)
910 			 *	return (1);
911 			 */
912 
913 			/*
914 			 * Setup un_orig_dev from device id info if the device
915 			 * is valid (not NODEV64).
916 			 */
917 			if (dev != NODEV64)
918 				column->un_orig_dev = dev;
919 
920 			if (column->un_devstate & RCS_RESYNC)
921 				resync_cnt++;
922 			if (column->un_devstate & (RCS_ERRED | RCS_LAST_ERRED))
923 				error_cnt++;
924 
925 			if (HOTSPARED(un, i)) {
926 				(void) md_hot_spare_ifc(HS_MKDEV,
927 				    0, 0, 0, &column->un_hs_id, NULL,
928 				    &hs, NULL);
929 				/*
930 				 * Same here
931 				 *
932 				 * if (hs == NODEV64)
933 				 *	return (1);
934 				 */
935 			}
936 
937 			if (HOTSPARED(un, i)) {
938 				if (column->un_devstate &
939 				    (RCS_OKAY | RCS_LAST_ERRED)) {
940 					column->un_dev = hs;
941 					column->un_pwstart =
942 					    column->un_hs_pwstart;
943 					column->un_devstart =
944 					    column->un_hs_devstart;
945 					preserve_flags &=
946 					    ~(MD_RAID_COPY_RESYNC |
947 					    MD_RAID_REGEN_RESYNC);
948 				} else  if (column->un_devstate & RCS_RESYNC) {
949 					/*
950 					 * if previous system was 4.0 set
951 					 * the direction flags
952 					 */
953 					if ((preserve_flags &
954 					    (MD_RAID_COPY_RESYNC |
955 					    MD_RAID_REGEN_RESYNC)) == 0) {
956 						if (column->un_alt_dev !=
957 						    NODEV64)
958 							preserve_flags |=
959 							    MD_RAID_COPY_RESYNC;
960 						else
961 							preserve_flags |=
962 							/* CSTYLED */
963 							    MD_RAID_REGEN_RESYNC;
964 					}
965 				}
966 			} else { /* no hot spares */
967 				column->un_dev = dev;
968 				column->un_pwstart = column->un_orig_pwstart;
969 				column->un_devstart = column->un_orig_devstart;
970 				if (column->un_devstate & RCS_RESYNC) {
971 					preserve_flags |= MD_RAID_REGEN_RESYNC;
972 					preserve_flags &= ~MD_RAID_COPY_RESYNC;
973 				}
974 			}
975 			if (! (column->un_devstate & RCS_RESYNC)) {
976 				preserve_flags &=
977 				    ~(MD_RAID_REGEN_RESYNC |
978 				    MD_RAID_COPY_RESYNC);
979 			}
980 
981 			column->un_devflags = preserve_flags;
982 			column->un_alt_dev = NODEV64;
983 			column->un_alt_pwstart = 0;
984 			column->un_alt_devstart = 0;
985 			un->un_resync_line_index = 0;
986 			un->un_resync_index = 0;
987 			un->un_percent_done = 0;
988 		}
989 	}
990 
991 	if (resync_cnt && error_cnt) {
992 		for (i = 0; i < un->un_totalcolumncnt; i++) {
993 			column  = &un->un_column[i];
994 			if (HOTSPARED(un, i) &&
995 			    (column->un_devstate & RCS_RESYNC) &&
996 			    (column->un_devflags & MD_RAID_COPY_RESYNC))
997 				/* hotspare has data */
998 				continue;
999 
1000 			if (HOTSPARED(un, i) &&
1001 			    (column->un_devstate & RCS_RESYNC)) {
1002 				/* hotspare does not have data */
1003 				raid_hs_release(HS_FREE, un, &hs_recid, i);
1004 				column->un_dev = column->un_orig_dev;
1005 				column->un_pwstart = column->un_orig_pwstart;
1006 				column->un_devstart = column->un_orig_devstart;
1007 				mddb_setrecprivate(hs_recid, MD_PRV_PENDCOM);
1008 			}
1009 
1010 			if (column->un_devstate & RCS_ERRED)
1011 				column->un_devstate = RCS_LAST_ERRED;
1012 
1013 			if (column->un_devstate & RCS_RESYNC)
1014 				column->un_devstate = RCS_ERRED;
1015 		}
1016 	}
1017 	mddb_setrecprivate(un->c.un_record_id, MD_PRV_PENDCOM);
1018 
1019 	un->un_pwid = 1; /* or some other possible value */
1020 	un->un_magic = RAID_UNMAGIC;
1021 	iosize = un->un_iosize;
1022 	un->un_pbuffer = kmem_alloc(dbtob(iosize), KM_SLEEP);
1023 	un->un_dbuffer = kmem_alloc(dbtob(iosize), KM_SLEEP);
1024 	mutex_init(&un->un_linlck_mx, NULL, MUTEX_DEFAULT, NULL);
1025 	cv_init(&un->un_linlck_cv, NULL, CV_DEFAULT, NULL);
1026 	un->un_linlck_chn = NULL;
1027 
1028 	/* place various information in the in-core data structures */
1029 	md_nblocks_set(mnum, un->c.un_total_blocks);
1030 	MD_UNIT(mnum) = un;
1031 
1032 	return (0);
1033 }
1034 
1035 /*
1036  * NAMES:	reset_raid
1037  * DESCRIPTION: RAID metadevice reset routine
1038  * PARAMETERS:	mr_unit_t *un - pointer to a unit structure
1039  *		minor_t mnum - RAID metadevice minor number
1040  *		int removing - a flag to imply removing device name from
1041  *			MDDB database.
1042  */
1043 void
1044 reset_raid(mr_unit_t *un, minor_t mnum, int removing)
1045 {
1046 	int		i, n = 0;
1047 	sv_dev_t	*sv;
1048 	mr_column_t	*column;
1049 	int		column_cnt = un->un_totalcolumncnt;
1050 	mddb_recid_t	*recids, vtoc_id;
1051 	int		hserr;
1052 
1053 	ASSERT((MDI_UNIT(mnum)->ui_io_lock->io_list_front == NULL) &&
1054 	    (MDI_UNIT(mnum)->ui_io_lock->io_list_back == NULL));
1055 
1056 	md_destroy_unit_incore(mnum, &raid_md_ops);
1057 
1058 	md_nblocks_set(mnum, -1ULL);
1059 	MD_UNIT(mnum) = NULL;
1060 
1061 	if (un->un_pbuffer) {
1062 		kmem_free(un->un_pbuffer, dbtob(un->un_iosize));
1063 		un->un_pbuffer = NULL;
1064 	}
1065 	if (un->un_dbuffer) {
1066 		kmem_free(un->un_dbuffer, dbtob(un->un_iosize));
1067 		un->un_dbuffer = NULL;
1068 	}
1069 
1070 	/* free all pre-write slots created during build incore */
1071 	for (i = 0; i < un->un_totalcolumncnt; i++)
1072 		raid_free_pw_reservation(un, i);
1073 
1074 	kmem_free(un->un_column_ic, sizeof (mr_column_ic_t) *
1075 	    un->un_totalcolumncnt);
1076 
1077 	kmem_free(un->mr_ic, sizeof (*un->mr_ic));
1078 
1079 	/*
1080 	 * Attempt release of its minor node
1081 	 */
1082 	md_remove_minor_node(mnum);
1083 
1084 	if (!removing)
1085 		return;
1086 
1087 	sv = (sv_dev_t *)kmem_zalloc((column_cnt + 1) * sizeof (sv_dev_t),
1088 	    KM_SLEEP);
1089 
1090 	recids = (mddb_recid_t *)
1091 	    kmem_zalloc((column_cnt + 2) * sizeof (mddb_recid_t), KM_SLEEP);
1092 
1093 	for (i = 0; i < column_cnt; i++) {
1094 		md_unit_t	*comp_un;
1095 		md_dev64_t	comp_dev;
1096 
1097 		column = &un->un_column[i];
1098 		sv[i].setno = MD_MIN2SET(mnum);
1099 		sv[i].key = column->un_orig_key;
1100 		if (HOTSPARED(un, i)) {
1101 			if (column->un_devstate & (RCS_ERRED | RCS_LAST_ERRED))
1102 				hserr = HS_BAD;
1103 			else
1104 				hserr = HS_FREE;
1105 			raid_hs_release(hserr, un, &recids[n++], i);
1106 		}
1107 		/*
1108 		 * deparent any metadevices.
1109 		 * NOTE: currently soft partitions are the only metadevices
1110 		 * allowed in RAID metadevices.
1111 		 */
1112 		comp_dev = column->un_dev;
1113 		if (md_getmajor(comp_dev) == md_major) {
1114 			comp_un = MD_UNIT(md_getminor(comp_dev));
1115 			recids[n++] = MD_RECID(comp_un);
1116 			md_reset_parent(comp_dev);
1117 		}
1118 	}
1119 	/* decrement the reference count of the old hsp */
1120 	if (un->un_hsp_id != -1)
1121 		(void) md_hot_spare_ifc(HSP_DECREF, un->un_hsp_id, 0, 0,
1122 		    &recids[n++], NULL, NULL, NULL);
1123 	recids[n] = 0;
1124 	MD_STATUS(un) |= MD_UN_BEING_RESET;
1125 	vtoc_id = un->c.un_vtoc_id;
1126 
1127 	raid_commit(un, recids);
1128 
1129 	/*
1130 	 * Remove self from the namespace
1131 	 */
1132 	if (un->c.un_revision & MD_FN_META_DEV) {
1133 		(void) md_rem_selfname(un->c.un_self_id);
1134 	}
1135 
1136 	/* Remove the unit structure */
1137 	mddb_deleterec_wrapper(un->c.un_record_id);
1138 
1139 	/* Remove the vtoc, if present */
1140 	if (vtoc_id)
1141 		mddb_deleterec_wrapper(vtoc_id);
1142 	md_rem_names(sv, column_cnt);
1143 	kmem_free(sv, (column_cnt + 1) * sizeof (sv_dev_t));
1144 	kmem_free(recids, (column_cnt + 2) * sizeof (mddb_recid_t));
1145 
1146 	SE_NOTIFY(EC_SVM_CONFIG, ESC_SVM_DELETE, SVM_TAG_METADEVICE,
1147 	    MD_MIN2SET(mnum), mnum);
1148 }
1149 
1150 /*
1151  * NAMES:	raid_error_parent
1152  * DESCRIPTION: mark a parent structure in error
1153  * PARAMETERS:	md_raidcs_t *cs - pointer to child structure
1154  *		int	error - error value to set
1155  * NOTE:	(TBR) - this routine currently is not in use.
1156  */
1157 static void
1158 raid_error_parent(md_raidps_t *ps, int error)
1159 {
1160 	mutex_enter(&ps->ps_mx);
1161 	ps->ps_flags |= MD_RPS_ERROR;
1162 	ps->ps_error = error;
1163 	mutex_exit(&ps->ps_mx);
1164 }
1165 
1166 /*
1167  * The following defines tell raid_free_parent
1168  *	RFP_RLS_LOCK		release the unit reader lock when done.
1169  *	RFP_DECR_PWFRAGS	decrement ps_pwfrags
1170  *	RFP_DECR_FRAGS		decrement ps_frags
1171  *	RFP_DECR_READFRAGS	read keeps FRAGS and PWFRAGS in lockstep
1172  */
1173 #define	RFP_RLS_LOCK		0x00001
1174 #define	RFP_DECR_PWFRAGS	0x00002
1175 #define	RFP_DECR_FRAGS		0x00004
1176 #define	RFP_DECR_READFRAGS	(RFP_DECR_PWFRAGS | RFP_DECR_FRAGS)
1177 
1178 /*
1179  * NAMES:	raid_free_parent
1180  * DESCRIPTION: free a parent structure
1181  * PARAMETERS:	md_raidcs_t *cs - pointer to child structure
1182  *		int	todo - indicates what needs to be done
1183  */
1184 static void
1185 raid_free_parent(md_raidps_t *ps, int todo)
1186 {
1187 	mdi_unit_t	*ui = ps->ps_ui;
1188 
1189 	ASSERT(ps->ps_magic == RAID_PSMAGIC);
1190 	ASSERT(ps->ps_flags & MD_RPS_INUSE);
1191 	mutex_enter(&ps->ps_mx);
1192 	if (todo & RFP_DECR_PWFRAGS) {
1193 		ASSERT(ps->ps_pwfrags);
1194 		ps->ps_pwfrags--;
1195 		if (ps->ps_pwfrags == 0 && (! (ps->ps_flags & MD_RPS_IODONE))) {
1196 			if (ps->ps_flags & MD_RPS_ERROR) {
1197 				ps->ps_bp->b_flags |= B_ERROR;
1198 				ps->ps_bp->b_error = ps->ps_error;
1199 			}
1200 			md_kstat_done(ui, ps->ps_bp, 0);
1201 			biodone(ps->ps_bp);
1202 			ps->ps_flags |= MD_RPS_IODONE;
1203 		}
1204 	}
1205 
1206 	if (todo & RFP_DECR_FRAGS) {
1207 		ASSERT(ps->ps_frags);
1208 		ps->ps_frags--;
1209 	}
1210 
1211 	if (ps->ps_frags != 0) {
1212 		mutex_exit(&ps->ps_mx);
1213 		return;
1214 	}
1215 
1216 	ASSERT((ps->ps_frags == 0) && (ps->ps_pwfrags == 0));
1217 	mutex_exit(&ps->ps_mx);
1218 
1219 	if (todo & RFP_RLS_LOCK)
1220 		md_io_readerexit(ui);
1221 
1222 	if (panicstr) {
1223 		ps->ps_flags |= MD_RPS_DONE;
1224 		return;
1225 	}
1226 
1227 	if (ps->ps_flags & MD_RPS_HSREQ)
1228 		(void) raid_hotspares();
1229 
1230 	ASSERT(todo & RFP_RLS_LOCK);
1231 	ps->ps_flags &= ~MD_RPS_INUSE;
1232 
1233 	md_dec_iocount(MD_MIN2SET(ps->ps_un->c.un_self_id));
1234 
1235 	kmem_cache_free(raid_parent_cache, ps);
1236 }
1237 
1238 /*
1239  * NAMES:	raid_free_child
1240  * DESCRIPTION: free a parent structure
1241  * PARAMETERS:	md_raidcs_t *cs - pointer to child structure
1242  *		int drop_locks	- 0 for no locks held
1243  * NOTE:	(TBR) - this routine currently is not in use.
1244  */
1245 static void
1246 raid_free_child(md_raidcs_t *cs, int drop_locks)
1247 {
1248 	mr_unit_t	*un = cs->cs_un;
1249 	md_raidcbuf_t	*cbuf, *cbuf1;
1250 
1251 	if (cs->cs_pw_inval_list)
1252 		raid_free_pwinvalidate(cs);
1253 
1254 	if (drop_locks) {
1255 		ASSERT(cs->cs_flags & MD_RCS_LLOCKD &&
1256 		    (cs->cs_flags & (MD_RCS_READER | MD_RCS_WRITER)));
1257 		md_unit_readerexit(MDI_UNIT(MD_SID(un)));
1258 		raid_line_exit(cs);
1259 	} else {
1260 		ASSERT(!(cs->cs_flags & MD_RCS_LLOCKD));
1261 	}
1262 
1263 	freebuffers(cs);
1264 	cbuf = cs->cs_buflist;
1265 	while (cbuf) {
1266 		cbuf1 = cbuf->cbuf_next;
1267 		kmem_cache_free(raid_cbuf_cache, cbuf);
1268 		cbuf = cbuf1;
1269 	}
1270 	if (cs->cs_dbuf.b_flags & B_REMAPPED)
1271 		bp_mapout(&cs->cs_dbuf);
1272 	kmem_cache_free(raid_child_cache, cs);
1273 }
1274 
1275 /*
1276  * NAME:	raid_regen_parity
1277  *
1278  * DESCRIPTION:	This routine is used to regenerate the parity blocks
1279  *		for the entire raid device.  It is called from
1280  *		both the regen thread and the IO path.
1281  *
1282  *		On error the entire device is marked as in error by
1283  *		placing the erroring device in error and all other
1284  *		devices in last_errored.
1285  *
1286  * PARAMETERS:	md_raidcs_t	*cs
1287  */
1288 void
1289 raid_regen_parity(md_raidcs_t *cs)
1290 {
1291 	mr_unit_t	*un = cs->cs_un;
1292 	mdi_unit_t	*ui = MDI_UNIT(un->c.un_self_id);
1293 	caddr_t		buffer;
1294 	caddr_t		parity_buffer;
1295 	buf_t		*bp;
1296 	uint_t		*dbuf, *pbuf;
1297 	uint_t		colcnt = un->un_totalcolumncnt;
1298 	int		column;
1299 	int		parity_column = cs->cs_pcolumn;
1300 	size_t		bcount;
1301 	int		j;
1302 
1303 	/*
1304 	 * This routine uses the data and parity buffers allocated to a
1305 	 * write.  In the case of a read the buffers are allocated and
1306 	 * freed at the end.
1307 	 */
1308 
1309 	ASSERT(IO_READER_HELD(un));
1310 	ASSERT(cs->cs_flags & MD_RCS_LLOCKD);
1311 	ASSERT(UNIT_READER_HELD(un));
1312 
1313 	if (raid_state_cnt(un, RCS_OKAY) != colcnt)
1314 		return;
1315 
1316 	if (cs->cs_flags & MD_RCS_READER) {
1317 		getpbuffer(cs);
1318 		getdbuffer(cs);
1319 	}
1320 	ASSERT(cs->cs_dbuffer && cs->cs_pbuffer);
1321 	bcount = cs->cs_bcount;
1322 	buffer = cs->cs_dbuffer;
1323 	parity_buffer = cs->cs_pbuffer;
1324 	bzero(parity_buffer, bcount);
1325 	bp = &cs->cs_dbuf;
1326 	for (column = 0; column < colcnt; column++) {
1327 		if (column == parity_column)
1328 			continue;
1329 		reset_buf(bp, B_READ | B_BUSY, bcount);
1330 		bp->b_un.b_addr = buffer;
1331 		bp->b_edev = md_dev64_to_dev(un->un_column[column].un_dev);
1332 		bp->b_lblkno = cs->cs_blkno + un->un_column[column].un_devstart;
1333 		bp->b_bcount = bcount;
1334 		bp->b_bufsize = bcount;
1335 		(void) md_call_strategy(bp, MD_STR_NOTTOP, NULL);
1336 		if (biowait(bp))
1337 			goto bail;
1338 		pbuf = (uint_t *)(void *)parity_buffer;
1339 		dbuf = (uint_t *)(void *)buffer;
1340 		for (j = 0; j < (bcount / (sizeof (uint_t))); j++) {
1341 			*pbuf = *pbuf ^ *dbuf;
1342 			pbuf++;
1343 			dbuf++;
1344 		}
1345 	}
1346 
1347 	reset_buf(bp, B_WRITE | B_BUSY, cs->cs_bcount);
1348 	bp->b_un.b_addr = parity_buffer;
1349 	bp->b_edev = md_dev64_to_dev(un->un_column[parity_column].un_dev);
1350 	bp->b_lblkno = cs->cs_blkno + un->un_column[parity_column].un_devstart;
1351 	bp->b_bcount = bcount;
1352 	bp->b_bufsize = bcount;
1353 	(void) md_call_strategy(bp, MD_STR_NOTTOP, NULL);
1354 	if (biowait(bp))
1355 		goto bail;
1356 
1357 	if (cs->cs_flags & MD_RCS_READER) {
1358 		freebuffers(cs);
1359 		cs->cs_pbuffer = NULL;
1360 		cs->cs_dbuffer = NULL;
1361 	}
1362 	bp->b_chain = (struct buf *)cs;
1363 	return;
1364 bail:
1365 	if (cs->cs_flags & MD_RCS_READER) {
1366 		freebuffers(cs);
1367 		cs->cs_pbuffer = NULL;
1368 		cs->cs_dbuffer = NULL;
1369 	}
1370 	md_unit_readerexit(ui);
1371 	un = md_unit_writerlock(ui);
1372 	raid_set_state(un, column, RCS_ERRED, 0);
1373 	for (column = 0; column < colcnt; column++)
1374 		raid_set_state(un, column, RCS_ERRED, 0);
1375 	raid_commit(un, NULL);
1376 	md_unit_writerexit(ui);
1377 	un = md_unit_readerlock(ui);
1378 	bp->b_chain = (struct buf *)cs;
1379 }
1380 
1381 /*
1382  * NAMES:	raid_error_state
1383  * DESCRIPTION: check unit and column states' impact on I/O error
1384  *		NOTE:	the state now may not be the state when the
1385  *			I/O completed due to race conditions.
1386  * PARAMETERS:	mr_unit_t *un - pointer to raid unit structure
1387  *		md_raidcs_t *cs - pointer to child structure
1388  *		buf_t	  *bp - pointer to buffer structure
1389  */
1390 static int
1391 raid_error_state(mr_unit_t *un, buf_t *bp)
1392 {
1393 	int		column;
1394 	int		i;
1395 
1396 	ASSERT(IO_READER_HELD(un));
1397 	ASSERT(UNIT_WRITER_HELD(un));
1398 
1399 	column = -1;
1400 	for (i = 0; i < un->un_totalcolumncnt; i++) {
1401 		if (un->un_column[i].un_dev == md_expldev(bp->b_edev)) {
1402 			column = i;
1403 			break;
1404 		}
1405 		if (un->un_column[i].un_alt_dev == md_expldev(bp->b_edev)) {
1406 			column = i;
1407 			break;
1408 		}
1409 	}
1410 
1411 	/* in case a replace snuck in while waiting on unit writer lock */
1412 
1413 	if (column == -1) {
1414 		return (0);
1415 	}
1416 
1417 	(void) raid_set_state(un, column, RCS_ERRED, 0);
1418 	ASSERT(un->un_state & (RUS_ERRED | RUS_LAST_ERRED));
1419 
1420 	raid_commit(un, NULL);
1421 	if (un->un_state & RUS_ERRED) {
1422 		SE_NOTIFY(EC_SVM_STATE, ESC_SVM_ERRED, SVM_TAG_METADEVICE,
1423 		    MD_UN2SET(un), MD_SID(un));
1424 	} else if (un->un_state & RUS_LAST_ERRED) {
1425 		SE_NOTIFY(EC_SVM_STATE, ESC_SVM_LASTERRED, SVM_TAG_METADEVICE,
1426 		    MD_UN2SET(un), MD_SID(un));
1427 	}
1428 
1429 	return (EIO);
1430 }
1431 
1432 /*
1433  * NAME:	raid_mapin_buf
1434  * DESCRIPTION:	wait for the input buffer header to be maped in
1435  * PARAMETERS:	md_raidps_t *ps
1436  */
1437 static void
1438 raid_mapin_buf(md_raidcs_t *cs)
1439 {
1440 	md_raidps_t	*ps = cs->cs_ps;
1441 
1442 	/*
1443 	 * check to see if the buffer is maped.  If all is ok return the
1444 	 * offset of the data and return.  Since it is expensive to grab
1445 	 * a mutex this is only done if the mapin is not complete.
1446 	 * Once the mutex is aquired it is possible that the mapin was
1447 	 * not done so recheck and if necessary do the mapin.
1448 	 */
1449 	if (ps->ps_mapin > 0) {
1450 		cs->cs_addr = ps->ps_addr + cs->cs_offset;
1451 		return;
1452 	}
1453 	mutex_enter(&ps->ps_mapin_mx);
1454 	if (ps->ps_mapin > 0) {
1455 		cs->cs_addr = ps->ps_addr + cs->cs_offset;
1456 		mutex_exit(&ps->ps_mapin_mx);
1457 		return;
1458 	}
1459 	bp_mapin(ps->ps_bp);
1460 	/*
1461 	 * get the new b_addr out of the parent since bp_mapin just changed it
1462 	 */
1463 	ps->ps_addr = ps->ps_bp->b_un.b_addr;
1464 	cs->cs_addr = ps->ps_addr + cs->cs_offset;
1465 	ps->ps_mapin++;
1466 	mutex_exit(&ps->ps_mapin_mx);
1467 }
1468 
1469 /*
1470  * NAMES:	raid_read_no_retry
1471  * DESCRIPTION: I/O retry routine for a RAID metadevice read
1472  *		read failed attempting to regenerate the data,
1473  *		no retry possible, error occured in raid_raidregenloop().
1474  * PARAMETERS:	mr_unit_t   *un - pointer to raid unit structure
1475  *		md_raidcs_t *cs - pointer to child structure
1476  */
1477 /*ARGSUSED*/
1478 static void
1479 raid_read_no_retry(mr_unit_t *un, md_raidcs_t *cs)
1480 {
1481 	md_raidps_t	*ps = cs->cs_ps;
1482 
1483 	raid_error_parent(ps, EIO);
1484 	raid_free_child(cs, 1);
1485 
1486 	/* decrement readfrags */
1487 	raid_free_parent(ps, RFP_DECR_READFRAGS | RFP_RLS_LOCK);
1488 }
1489 
1490 /*
1491  * NAMES:	raid_read_retry
1492  * DESCRIPTION: I/O retry routine for a RAID metadevice read
1493  * PARAMETERS:	md_raidcs_t *cs - pointer to child structure
1494  */
1495 static void
1496 raid_read_retry(mr_unit_t *un, md_raidcs_t *cs)
1497 {
1498 	/* re-initialize the buf_t structure for raid_read() */
1499 	cs->cs_dbuf.b_chain = (struct buf *)cs;
1500 	cs->cs_dbuf.b_back = &cs->cs_dbuf;
1501 	cs->cs_dbuf.b_forw = &cs->cs_dbuf;
1502 	cs->cs_dbuf.b_flags = B_BUSY;	/* initialize flags */
1503 	cs->cs_dbuf.b_error = 0;	/* initialize error */
1504 	cs->cs_dbuf.b_offset = -1;
1505 	/* Initialize semaphores */
1506 	sema_init(&cs->cs_dbuf.b_io, 0, NULL,
1507 	    SEMA_DEFAULT, NULL);
1508 	sema_init(&cs->cs_dbuf.b_sem, 0, NULL,
1509 	    SEMA_DEFAULT, NULL);
1510 
1511 	cs->cs_pbuf.b_chain = (struct buf *)cs;
1512 	cs->cs_pbuf.b_back = &cs->cs_pbuf;
1513 	cs->cs_pbuf.b_forw = &cs->cs_pbuf;
1514 	cs->cs_pbuf.b_flags = B_BUSY;	/* initialize flags */
1515 	cs->cs_pbuf.b_error = 0;	/* initialize error */
1516 	cs->cs_pbuf.b_offset = -1;
1517 	sema_init(&cs->cs_pbuf.b_io, 0, NULL,
1518 	    SEMA_DEFAULT, NULL);
1519 	sema_init(&cs->cs_pbuf.b_sem, 0, NULL,
1520 	    SEMA_DEFAULT, NULL);
1521 
1522 	cs->cs_flags &= ~MD_RCS_ERROR;	/* reset child error flag */
1523 	cs->cs_flags |= MD_RCS_RECOVERY;  /* set RECOVERY flag */
1524 
1525 	/*
1526 	 * re-scheduling I/O with raid_read_io() is simpler. basically,
1527 	 * raid_read_io() is invoked again with same child structure.
1528 	 * (NOTE: we aren`t supposed to do any error recovery when an I/O
1529 	 * error occured in raid_raidregenloop().
1530 	 */
1531 	raid_mapin_buf(cs);
1532 	raid_read_io(un, cs);
1533 }
1534 
1535 /*
1536  * NAMES:	raid_rderr
1537  * DESCRIPTION: I/O error handling routine for a RAID metadevice read
1538  * PARAMETERS:	md_raidcs_t *cs - pointer to child structure
1539  * LOCKS:	must obtain unit writer lock while calling raid_error_state
1540  *		since a unit or column state transition may take place.
1541  *		must obtain unit reader lock to retry I/O.
1542  */
1543 /*ARGSUSED*/
1544 static void
1545 raid_rderr(md_raidcs_t *cs)
1546 {
1547 	md_raidps_t	*ps;
1548 	mdi_unit_t	*ui;
1549 	mr_unit_t	*un;
1550 	int		error = 0;
1551 
1552 	ps = cs->cs_ps;
1553 	ui = ps->ps_ui;
1554 	un = (mr_unit_t *)md_unit_writerlock(ui);
1555 	ASSERT(un != 0);
1556 
1557 	if (cs->cs_dbuf.b_flags & B_ERROR)
1558 		error = raid_error_state(un, &cs->cs_dbuf);
1559 	if (cs->cs_pbuf.b_flags & B_ERROR)
1560 		error |= raid_error_state(un, &cs->cs_pbuf);
1561 
1562 	md_unit_writerexit(ui);
1563 
1564 	ps->ps_flags |= MD_RPS_HSREQ;
1565 
1566 	un = (mr_unit_t *)md_unit_readerlock(ui);
1567 	ASSERT(un != 0);
1568 	/* now attempt the appropriate retry routine */
1569 	(*(cs->cs_retry_call))(un, cs);
1570 }
1571 
1572 
1573 /*
1574  * NAMES:	raid_read_error
1575  * DESCRIPTION: I/O error handling routine for a RAID metadevice read
1576  * PARAMETERS:	md_raidcs_t *cs - pointer to child structure
1577  */
1578 /*ARGSUSED*/
1579 static void
1580 raid_read_error(md_raidcs_t *cs)
1581 {
1582 	md_raidps_t	*ps;
1583 	mdi_unit_t	*ui;
1584 	mr_unit_t	*un;
1585 	set_t		setno;
1586 
1587 	ps = cs->cs_ps;
1588 	ui = ps->ps_ui;
1589 	un = cs->cs_un;
1590 
1591 	setno = MD_UN2SET(un);
1592 
1593 	if ((cs->cs_dbuf.b_flags & B_ERROR) &&
1594 	    (COLUMN_STATE(un, cs->cs_dcolumn) != RCS_ERRED) &&
1595 	    (COLUMN_STATE(un, cs->cs_dcolumn) != RCS_LAST_ERRED))
1596 		cmn_err(CE_WARN, "md %s: read error on %s",
1597 		    md_shortname(MD_SID(un)),
1598 		    md_devname(setno, md_expldev(cs->cs_dbuf.b_edev), NULL, 0));
1599 
1600 	if ((cs->cs_pbuf.b_flags & B_ERROR) &&
1601 	    (COLUMN_STATE(un, cs->cs_pcolumn) != RCS_ERRED) &&
1602 	    (COLUMN_STATE(un, cs->cs_pcolumn) != RCS_LAST_ERRED))
1603 		cmn_err(CE_WARN, "md %s: read error on %s",
1604 		    md_shortname(MD_SID(un)),
1605 		    md_devname(setno, md_expldev(cs->cs_pbuf.b_edev), NULL, 0));
1606 
1607 	md_unit_readerexit(ui);
1608 
1609 	ASSERT(cs->cs_frags == 0);
1610 
1611 	/* now schedule processing for possible state change */
1612 	daemon_request(&md_mstr_daemon, raid_rderr,
1613 	    (daemon_queue_t *)cs, REQ_OLD);
1614 
1615 }
1616 
1617 /*
1618  * NAMES:	getdbuffer
1619  * DESCRIPTION: data buffer allocation for a child structure
1620  * PARAMETERS:	md_raidcs_t *cs - pointer to child structure
1621  *
1622  * NOTE: always get dbuffer before pbuffer
1623  *	 and get both buffers before pwslot
1624  *	 otherwise a deadlock could be introduced.
1625  */
1626 static void
1627 getdbuffer(md_raidcs_t *cs)
1628 {
1629 	mr_unit_t	*un;
1630 
1631 	cs->cs_dbuffer = kmem_alloc(cs->cs_bcount + DEV_BSIZE, KM_NOSLEEP);
1632 	if (cs->cs_dbuffer != NULL)
1633 		return;
1634 	un = cs->cs_ps->ps_un;
1635 	mutex_enter(&un->un_mx);
1636 	while (un->un_dbuffer == NULL) {
1637 		STAT_INC(data_buffer_waits);
1638 		un->un_rflags |= MD_RFLAG_NEEDBUF;
1639 		cv_wait(&un->un_cv, &un->un_mx);
1640 	}
1641 	cs->cs_dbuffer = un->un_dbuffer;
1642 	cs->cs_flags |= MD_RCS_UNDBUF;
1643 	un->un_dbuffer = NULL;
1644 	mutex_exit(&un->un_mx);
1645 }
1646 
1647 /*
1648  * NAMES:	getpbuffer
1649  * DESCRIPTION: parity buffer allocation for a child structure
1650  * PARAMETERS:	md_raidcs_t *cs - pointer to child structure
1651  *
1652  * NOTE: always get dbuffer before pbuffer
1653  *	 and get both buffers before pwslot
1654  *	 otherwise a deadlock could be introduced.
1655  */
1656 static void
1657 getpbuffer(md_raidcs_t *cs)
1658 {
1659 	mr_unit_t *un;
1660 
1661 	cs->cs_pbuffer = kmem_alloc(cs->cs_bcount + DEV_BSIZE, KM_NOSLEEP);
1662 	if (cs->cs_pbuffer != NULL)
1663 		return;
1664 	un = cs->cs_ps->ps_un;
1665 	mutex_enter(&un->un_mx);
1666 	while (un->un_pbuffer == NULL) {
1667 		STAT_INC(parity_buffer_waits);
1668 		un->un_rflags |= MD_RFLAG_NEEDBUF;
1669 		cv_wait(&un->un_cv, &un->un_mx);
1670 	}
1671 	cs->cs_pbuffer = un->un_pbuffer;
1672 	cs->cs_flags |= MD_RCS_UNPBUF;
1673 	un->un_pbuffer = NULL;
1674 	mutex_exit(&un->un_mx);
1675 }
1676 static void
1677 getresources(md_raidcs_t *cs)
1678 {
1679 	md_raidcbuf_t	*cbuf;
1680 	/*
1681 	 * NOTE: always get dbuffer before pbuffer
1682 	 *	 and get both buffers before pwslot
1683 	 *	 otherwise a deadlock could be introduced.
1684 	 */
1685 	getdbuffer(cs);
1686 	getpbuffer(cs);
1687 	for (cbuf = cs->cs_buflist; cbuf; cbuf = cbuf->cbuf_next)
1688 		cbuf->cbuf_buffer =
1689 		    kmem_alloc(cs->cs_bcount + DEV_BSIZE, KM_SLEEP);
1690 }
1691 /*
1692  * NAMES:	freebuffers
1693  * DESCRIPTION: child structure buffer freeing routine
1694  * PARAMETERS:	md_raidcs_t *cs - pointer to child structure
1695  */
1696 static void
1697 freebuffers(md_raidcs_t *cs)
1698 {
1699 	mr_unit_t	*un;
1700 	md_raidcbuf_t	*cbuf;
1701 
1702 	/* free buffers used for full line write */
1703 	for (cbuf = cs->cs_buflist; cbuf; cbuf = cbuf->cbuf_next) {
1704 		if (cbuf->cbuf_buffer == NULL)
1705 			continue;
1706 		kmem_free(cbuf->cbuf_buffer, cbuf->cbuf_bcount + DEV_BSIZE);
1707 		cbuf->cbuf_buffer = NULL;
1708 		cbuf->cbuf_bcount = 0;
1709 	}
1710 
1711 	if (cs->cs_flags & (MD_RCS_UNDBUF | MD_RCS_UNPBUF)) {
1712 		un = cs->cs_un;
1713 		mutex_enter(&un->un_mx);
1714 	}
1715 	if (cs->cs_dbuffer) {
1716 		if (cs->cs_flags & MD_RCS_UNDBUF)
1717 			un->un_dbuffer = cs->cs_dbuffer;
1718 		else
1719 			kmem_free(cs->cs_dbuffer, cs->cs_bcount + DEV_BSIZE);
1720 	}
1721 	if (cs->cs_pbuffer) {
1722 		if (cs->cs_flags & MD_RCS_UNPBUF)
1723 			un->un_pbuffer = cs->cs_pbuffer;
1724 		else
1725 			kmem_free(cs->cs_pbuffer, cs->cs_bcount + DEV_BSIZE);
1726 	}
1727 	if (cs->cs_flags & (MD_RCS_UNDBUF | MD_RCS_UNPBUF)) {
1728 		un->un_rflags &= ~MD_RFLAG_NEEDBUF;
1729 		cv_broadcast(&un->un_cv);
1730 		mutex_exit(&un->un_mx);
1731 	}
1732 }
1733 
1734 /*
1735  * NAMES:	raid_line_reader_lock, raid_line_writer_lock
1736  * DESCRIPTION: RAID metadevice line reader and writer lock routines
1737  *		data column # and parity column #.
1738  * PARAMETERS:	md_raidcs_t *cs - pointer to child structure
1739  */
1740 
1741 void
1742 raid_line_reader_lock(md_raidcs_t *cs, int resync_thread)
1743 {
1744 	mr_unit_t	*un;
1745 	md_raidcs_t	*cs1;
1746 
1747 	ASSERT(cs->cs_line != MD_DISKADDR_ERROR);
1748 	un = cs->cs_un;
1749 	cs->cs_flags |= MD_RCS_READER;
1750 	STAT_CHECK(raid_line_lock_wait, MUTEX_HELD(&un->un_linlck_mx));
1751 	if (!panicstr)
1752 		mutex_enter(&un->un_linlck_mx);
1753 	cs1 = un->un_linlck_chn;
1754 	while (cs1 != NULL) {
1755 		for (cs1 = un->un_linlck_chn; cs1; cs1 = cs1->cs_linlck_next)
1756 			if (raid_io_overlaps(cs, cs1) == 1)
1757 				if (cs1->cs_flags & MD_RCS_WRITER)
1758 					break;
1759 
1760 		if (cs1 != NULL) {
1761 			if (panicstr)
1762 				panic("md; raid line write lock held");
1763 			un->un_linlck_flg = 1;
1764 			cv_wait(&un->un_linlck_cv, &un->un_linlck_mx);
1765 			STAT_INC(raid_read_waits);
1766 		}
1767 	}
1768 	STAT_MAX(raid_max_reader_locks, raid_reader_locks_active);
1769 	STAT_INC(raid_reader_locks);
1770 	cs1 = un->un_linlck_chn;
1771 	if (cs1 != NULL)
1772 		cs1->cs_linlck_prev = cs;
1773 	cs->cs_linlck_next = cs1;
1774 	cs->cs_linlck_prev = NULL;
1775 	un->un_linlck_chn = cs;
1776 	cs->cs_flags |= MD_RCS_LLOCKD;
1777 	if (resync_thread) {
1778 		diskaddr_t lastblk = cs->cs_blkno + cs->cs_blkcnt - 1;
1779 		diskaddr_t line = (lastblk + 1) / un->un_segsize;
1780 		ASSERT(raid_state_cnt(un, RCS_RESYNC));
1781 		mutex_enter(&un->un_mx);
1782 		un->un_resync_line_index = line;
1783 		mutex_exit(&un->un_mx);
1784 	}
1785 	if (!panicstr)
1786 		mutex_exit(&un->un_linlck_mx);
1787 }
1788 
1789 int
1790 raid_line_writer_lock(md_raidcs_t *cs, int lock)
1791 {
1792 	mr_unit_t	*un;
1793 	md_raidcs_t	*cs1;
1794 
1795 	ASSERT(cs->cs_line != MD_DISKADDR_ERROR);
1796 	cs->cs_flags |= MD_RCS_WRITER;
1797 	un = cs->cs_ps->ps_un;
1798 
1799 	STAT_CHECK(raid_line_lock_wait, MUTEX_HELD(&un->un_linlck_mx));
1800 	if (lock && !panicstr)
1801 		mutex_enter(&un->un_linlck_mx);
1802 	ASSERT(MUTEX_HELD(&un->un_linlck_mx));
1803 
1804 	cs1 = un->un_linlck_chn;
1805 	for (cs1 = un->un_linlck_chn; cs1; cs1 = cs1->cs_linlck_next)
1806 		if (raid_io_overlaps(cs, cs1))
1807 			break;
1808 
1809 	if (cs1 != NULL) {
1810 		if (panicstr)
1811 			panic("md: line writer lock inaccessible");
1812 		goto no_lock_exit;
1813 	}
1814 
1815 	if (raid_alloc_pwslot(cs)) {
1816 		if (panicstr)
1817 			panic("md: no prewrite slots");
1818 		STAT_INC(raid_prewrite_waits);
1819 		goto no_lock_exit;
1820 	}
1821 
1822 	cs1 = un->un_linlck_chn;
1823 	if (cs1 != NULL)
1824 		cs1->cs_linlck_prev = cs;
1825 	cs->cs_linlck_next = cs1;
1826 	cs->cs_linlck_prev = NULL;
1827 	un->un_linlck_chn = cs;
1828 	cs->cs_flags |= MD_RCS_LLOCKD;
1829 	cs->cs_flags &= ~MD_RCS_WAITING;
1830 	STAT_INC(raid_writer_locks);
1831 	STAT_MAX(raid_max_write_locks, raid_write_locks_active);
1832 	if (lock && !panicstr)
1833 		mutex_exit(&un->un_linlck_mx);
1834 	return (0);
1835 
1836 no_lock_exit:
1837 	/* if this is already queued then do not requeue it */
1838 	ASSERT(! (cs->cs_flags & MD_RCS_LLOCKD));
1839 	if (!lock || (cs->cs_flags & MD_RCS_WAITING))
1840 		return (1);
1841 	cs->cs_flags |= MD_RCS_WAITING;
1842 	cs->cs_un = un;
1843 	raid_enqueue(cs);
1844 	if (lock && !panicstr)
1845 		mutex_exit(&un->un_linlck_mx);
1846 	return (1);
1847 }
1848 
1849 static void
1850 raid_startio(md_raidcs_t *cs)
1851 {
1852 	mdi_unit_t	*ui = cs->cs_ps->ps_ui;
1853 	mr_unit_t	*un = cs->cs_un;
1854 
1855 	un = md_unit_readerlock(ui);
1856 	raid_write_io(un, cs);
1857 }
1858 
1859 void
1860 raid_io_startup(mr_unit_t *un)
1861 {
1862 	md_raidcs_t	*waiting_list, *cs1;
1863 	md_raidcs_t	*previous = NULL, *next = NULL;
1864 	mdi_unit_t	*ui =  MDI_UNIT(un->c.un_self_id);
1865 	kmutex_t	*io_list_mutex = &ui->ui_io_lock->io_list_mutex;
1866 
1867 	ASSERT(MUTEX_HELD(&un->un_linlck_mx));
1868 	mutex_enter(io_list_mutex);
1869 
1870 	/*
1871 	 * check to be sure there are no reader locks outstanding.  If
1872 	 * there are not then pass on the writer lock.
1873 	 */
1874 	waiting_list = ui->ui_io_lock->io_list_front;
1875 	while (waiting_list) {
1876 		ASSERT(waiting_list->cs_flags & MD_RCS_WAITING);
1877 		ASSERT(! (waiting_list->cs_flags & MD_RCS_LLOCKD));
1878 		for (cs1 = un->un_linlck_chn; cs1; cs1 = cs1->cs_linlck_next)
1879 			if (raid_io_overlaps(waiting_list, cs1) == 1)
1880 				break;
1881 		/*
1882 		 * there was an IOs that overlaps this io so go onto
1883 		 * the next io in the waiting list
1884 		 */
1885 		if (cs1) {
1886 			previous = waiting_list;
1887 			waiting_list = waiting_list->cs_linlck_next;
1888 			continue;
1889 		}
1890 
1891 		/*
1892 		 * There are no IOs that overlap this, so remove it from
1893 		 * the waiting queue, and start it
1894 		 */
1895 
1896 		if (raid_check_pw(waiting_list)) {
1897 			ASSERT(waiting_list->cs_flags & MD_RCS_WAITING);
1898 			previous = waiting_list;
1899 			waiting_list = waiting_list->cs_linlck_next;
1900 			continue;
1901 		}
1902 		ASSERT(waiting_list->cs_flags & MD_RCS_WAITING);
1903 
1904 		next = waiting_list->cs_linlck_next;
1905 		if (previous)
1906 			previous->cs_linlck_next = next;
1907 		else
1908 			ui->ui_io_lock->io_list_front = next;
1909 
1910 		if (ui->ui_io_lock->io_list_front == NULL)
1911 			ui->ui_io_lock->io_list_back = NULL;
1912 
1913 		if (ui->ui_io_lock->io_list_back == waiting_list)
1914 			ui->ui_io_lock->io_list_back = previous;
1915 
1916 		waiting_list->cs_linlck_next = NULL;
1917 		waiting_list->cs_flags &= ~MD_RCS_WAITING;
1918 		STAT_DEC(raid_write_queue_length);
1919 		if (raid_line_writer_lock(waiting_list, 0))
1920 			panic("region locking corrupted");
1921 
1922 		ASSERT(waiting_list->cs_flags & MD_RCS_LLOCKD);
1923 		daemon_request(&md_mstr_daemon, raid_startio,
1924 		    (daemon_queue_t *)waiting_list, REQ_OLD);
1925 		waiting_list = next;
1926 
1927 	}
1928 	mutex_exit(io_list_mutex);
1929 }
1930 
1931 void
1932 raid_line_exit(md_raidcs_t *cs)
1933 {
1934 	mr_unit_t	*un;
1935 
1936 	un = cs->cs_ps->ps_un;
1937 	STAT_CHECK(raid_line_lock_wait, MUTEX_HELD(&un->un_linlck_mx));
1938 	mutex_enter(&un->un_linlck_mx);
1939 	if (cs->cs_flags & MD_RCS_READER)
1940 		STAT_DEC(raid_reader_locks_active);
1941 	else
1942 		STAT_DEC(raid_write_locks_active);
1943 
1944 	if (cs->cs_linlck_prev)
1945 		cs->cs_linlck_prev->cs_linlck_next = cs->cs_linlck_next;
1946 	else
1947 		un->un_linlck_chn = cs->cs_linlck_next;
1948 	if (cs->cs_linlck_next)
1949 		cs->cs_linlck_next->cs_linlck_prev = cs->cs_linlck_prev;
1950 
1951 	cs->cs_flags &= ~MD_RCS_LLOCKD;
1952 
1953 	if (un->un_linlck_flg)
1954 		cv_broadcast(&un->un_linlck_cv);
1955 
1956 	un->un_linlck_flg = 0;
1957 	cs->cs_line = MD_DISKADDR_ERROR;
1958 
1959 	raid_cancel_pwslot(cs);
1960 	/*
1961 	 * now that the lock is droped go ahead and see if there are any
1962 	 * other writes that can be started up
1963 	 */
1964 	raid_io_startup(un);
1965 
1966 	mutex_exit(&un->un_linlck_mx);
1967 }
1968 
1969 /*
1970  * NAMES:	raid_line, raid_pcolumn, raid_dcolumn
1971  * DESCRIPTION: RAID metadevice APIs for mapping segment # to line #,
1972  *		data column # and parity column #.
1973  * PARAMETERS:	int segment - segment number
1974  *		mr_unit_t *un - pointer to an unit structure
1975  * RETURNS:	raid_line returns line #
1976  *		raid_dcolumn returns data column #
1977  *		raid_pcolumn returns parity column #
1978  */
1979 static diskaddr_t
1980 raid_line(diskaddr_t segment, mr_unit_t *un)
1981 {
1982 	diskaddr_t	adj_seg;
1983 	diskaddr_t	line;
1984 	diskaddr_t	max_orig_segment;
1985 
1986 	max_orig_segment = (un->un_origcolumncnt - 1) * un->un_segsincolumn;
1987 	if (segment >= max_orig_segment) {
1988 		adj_seg = segment - max_orig_segment;
1989 		line = adj_seg % un->un_segsincolumn;
1990 	} else {
1991 		line = segment / (un->un_origcolumncnt - 1);
1992 	}
1993 	return (line);
1994 }
1995 
1996 uint_t
1997 raid_dcolumn(diskaddr_t segment, mr_unit_t *un)
1998 {
1999 	diskaddr_t	adj_seg;
2000 	diskaddr_t	line;
2001 	diskaddr_t	max_orig_segment;
2002 	uint_t		column;
2003 
2004 	max_orig_segment = (un->un_origcolumncnt - 1) * un->un_segsincolumn;
2005 	if (segment >= max_orig_segment) {
2006 		adj_seg = segment - max_orig_segment;
2007 		column = un->un_origcolumncnt  +
2008 		    (uint_t)(adj_seg / un->un_segsincolumn);
2009 	} else {
2010 		line = segment / (un->un_origcolumncnt - 1);
2011 		column = (uint_t)((segment %
2012 		    (un->un_origcolumncnt - 1) + line) % un->un_origcolumncnt);
2013 	}
2014 	return (column);
2015 }
2016 
2017 uint_t
2018 raid_pcolumn(diskaddr_t segment, mr_unit_t *un)
2019 {
2020 	diskaddr_t	adj_seg;
2021 	diskaddr_t	line;
2022 	diskaddr_t	max_orig_segment;
2023 	uint_t		column;
2024 
2025 	max_orig_segment = (un->un_origcolumncnt - 1) * un->un_segsincolumn;
2026 	if (segment >= max_orig_segment) {
2027 		adj_seg = segment - max_orig_segment;
2028 		line = adj_seg % un->un_segsincolumn;
2029 	} else {
2030 		line = segment / (un->un_origcolumncnt - 1);
2031 	}
2032 	column = (uint_t)((line + (un->un_origcolumncnt - 1)) %
2033 	    un->un_origcolumncnt);
2034 	return (column);
2035 }
2036 
2037 
2038 /*
2039  * Is called in raid_iosetup to probe each column to insure
2040  * that all the columns are in 'okay' state and meet the
2041  * 'full line' requirement.  If any column is in error,
2042  * we don't want to enable the 'full line' flag.  Previously,
2043  * we would do so and disable it only when a error is
2044  * detected after the first 'full line' io which is too late
2045  * and leads to the potential data corruption.
2046  */
2047 static int
2048 raid_check_cols(mr_unit_t *un)
2049 {
2050 	buf_t		bp;
2051 	char		*buf;
2052 	mr_column_t	*colptr;
2053 	minor_t		mnum = MD_SID(un);
2054 	int		i;
2055 	int		err = 0;
2056 
2057 	buf = kmem_zalloc((uint_t)DEV_BSIZE, KM_SLEEP);
2058 
2059 	for (i = 0; i < un->un_totalcolumncnt; i++) {
2060 		md_dev64_t tmpdev;
2061 
2062 		colptr = &un->un_column[i];
2063 
2064 		tmpdev = colptr->un_dev;
2065 		/*
2066 		 * Open by device id
2067 		 * If this device is hotspared
2068 		 * use the hotspare key
2069 		 */
2070 		tmpdev = md_resolve_bydevid(mnum, tmpdev, HOTSPARED(un, i) ?
2071 		    colptr->un_hs_key : colptr->un_orig_key);
2072 
2073 		if (tmpdev == NODEV64) {
2074 			err = 1;
2075 			break;
2076 		}
2077 
2078 		colptr->un_dev = tmpdev;
2079 
2080 		bzero((caddr_t)&bp, sizeof (buf_t));
2081 		bp.b_back = &bp;
2082 		bp.b_forw = &bp;
2083 		bp.b_flags = (B_READ | B_BUSY);
2084 		sema_init(&bp.b_io, 0, NULL,
2085 		    SEMA_DEFAULT, NULL);
2086 		sema_init(&bp.b_sem, 0, NULL,
2087 		    SEMA_DEFAULT, NULL);
2088 		bp.b_edev = md_dev64_to_dev(colptr->un_dev);
2089 		bp.b_lblkno = colptr->un_pwstart;
2090 		bp.b_bcount = DEV_BSIZE;
2091 		bp.b_bufsize = DEV_BSIZE;
2092 		bp.b_un.b_addr = (caddr_t)buf;
2093 		(void) md_call_strategy(&bp, 0, NULL);
2094 		if (biowait(&bp)) {
2095 			err = 1;
2096 			break;
2097 		}
2098 	}
2099 
2100 	kmem_free(buf, DEV_BSIZE);
2101 	return (err);
2102 }
2103 
2104 /*
2105  * NAME:	raid_iosetup
2106  * DESCRIPTION: RAID metadevice specific I/O set up routine which does
2107  *		all the necessary calculations to determine the location
2108  *		of the segement for the I/O.
2109  * PARAMETERS:	mr_unit_t *un - unit number of RAID metadevice
2110  *		diskaddr_t	blkno - block number of the I/O attempt
2111  *		size_t		blkcnt - block count for this I/O
2112  *		md_raidcs_t *cs - child structure for each segmented I/O
2113  *
2114  * NOTE:	The following is an example of a raid disk layer out:
2115  *
2116  *		Total Column = 5
2117  *		Original Column = 4
2118  *		Segment Per Column = 10
2119  *
2120  *			Col#0	Col#1	Col#2	Col#3	Col#4	Col#5	Col#6
2121  *		-------------------------------------------------------------
2122  *		line#0	Seg#0	Seg#1	Seg#2	Parity	Seg#30	Seg#40
2123  *		line#1	Parity	Seg#3	Seg#4	Seg#5	Seg#31
2124  *		line#2	Seg#8	Parity	Seg#6	Seg#7	Seg#32
2125  *		line#3	Seg#10	Seg#11	Parity	Seg#9	Seg#33
2126  *		line#4	Seg#12	Seg#13	Seg#14	Parity	Seg#34
2127  *		line#5	Parity	Seg#15	Seg#16	Seg#17	Seg#35
2128  *		line#6	Seg#20	Parity	Seg#18	Seg#19	Seg#36
2129  *		line#7	Seg#22	Seg#23	Parity	Seg#21	Seg#37
2130  *		line#8	Seg#24	Seg#25	Seg#26	Parity	Seg#38
2131  *		line#9	Parity	Seg#27	Seg#28	Seg#29	Seg#39
2132  */
2133 static size_t
2134 raid_iosetup(
2135 	mr_unit_t	*un,
2136 	diskaddr_t	blkno,
2137 	size_t		blkcnt,
2138 	md_raidcs_t	*cs
2139 )
2140 {
2141 	diskaddr_t	segment;
2142 	diskaddr_t	segstart;
2143 	diskaddr_t	segoff;
2144 	size_t		leftover;
2145 	diskaddr_t	line;
2146 	uint_t		iosize;
2147 	uint_t		colcnt;
2148 
2149 	/* caculate the segment# and offset for the block */
2150 	segment = blkno / un->un_segsize;
2151 	segstart = segment * un->un_segsize;
2152 	segoff = blkno - segstart;
2153 	iosize = un->un_iosize - 1;
2154 	colcnt = un->un_totalcolumncnt - 1;
2155 	line = raid_line(segment, un);
2156 	cs->cs_dcolumn = raid_dcolumn(segment, un);
2157 	cs->cs_pcolumn = raid_pcolumn(segment, un);
2158 	cs->cs_dflags = un->un_column[cs->cs_dcolumn].un_devflags;
2159 	cs->cs_pflags = un->un_column[cs->cs_pcolumn].un_devflags;
2160 	cs->cs_line = line;
2161 
2162 	if ((cs->cs_ps->ps_flags & MD_RPS_WRITE) &&
2163 	    (UNIT_STATE(un) & RCS_OKAY) &&
2164 	    (segoff == 0) &&
2165 	    (un->un_totalcolumncnt == un->un_origcolumncnt) &&
2166 	    (un->un_segsize < un->un_iosize) &&
2167 	    (un->un_iosize <= un->un_maxio) &&
2168 	    (blkno == line * un->un_segsize * colcnt) &&
2169 	    (blkcnt >= ((un->un_totalcolumncnt -1) * un->un_segsize)) &&
2170 	    (raid_state_cnt(un, RCS_OKAY) == un->un_origcolumncnt) &&
2171 	    (raid_check_cols(un) == 0)) {
2172 
2173 		md_raidcbuf_t	**cbufp;
2174 		md_raidcbuf_t	*cbuf;
2175 		int		i, j;
2176 
2177 		STAT_INC(raid_full_line_writes);
2178 		leftover = blkcnt - (un->un_segsize * colcnt);
2179 		ASSERT(blkcnt >= (un->un_segsize * colcnt));
2180 		cs->cs_blkno = line * un->un_segsize;
2181 		cs->cs_blkcnt = un->un_segsize;
2182 		cs->cs_lastblk = cs->cs_blkno + cs->cs_blkcnt - 1;
2183 		cs->cs_bcount = dbtob(cs->cs_blkcnt);
2184 		cs->cs_flags |= MD_RCS_LINE;
2185 
2186 		cbufp = &cs->cs_buflist;
2187 		for (i = 0; i < un->un_totalcolumncnt; i++) {
2188 			j = cs->cs_dcolumn + i;
2189 			j = j % un->un_totalcolumncnt;
2190 
2191 			if ((j == cs->cs_dcolumn) || (j == cs->cs_pcolumn))
2192 				continue;
2193 			cbuf = kmem_cache_alloc(raid_cbuf_cache,
2194 			    MD_ALLOCFLAGS);
2195 			raid_cbuf_init(cbuf);
2196 			cbuf->cbuf_un = cs->cs_un;
2197 			cbuf->cbuf_ps = cs->cs_ps;
2198 			cbuf->cbuf_column = j;
2199 			cbuf->cbuf_bcount = dbtob(un->un_segsize);
2200 			*cbufp = cbuf;
2201 			cbufp = &cbuf->cbuf_next;
2202 		}
2203 		return (leftover);
2204 	}
2205 
2206 	leftover = blkcnt - (un->un_segsize - segoff);
2207 	if (blkcnt > (un->un_segsize - segoff))
2208 		blkcnt -= leftover;
2209 	else
2210 		leftover = 0;
2211 
2212 	if (blkcnt > (size_t)iosize) {
2213 		leftover += (blkcnt - iosize);
2214 		blkcnt = iosize;
2215 	}
2216 
2217 	/* calculate the line# and column# for the segment */
2218 	cs->cs_flags &= ~MD_RCS_LINE;
2219 	cs->cs_blkno = line * un->un_segsize + segoff;
2220 	cs->cs_blkcnt = (uint_t)blkcnt;
2221 	cs->cs_lastblk = cs->cs_blkno + cs->cs_blkcnt - 1;
2222 	cs->cs_bcount = dbtob((uint_t)blkcnt);
2223 	return (leftover);
2224 }
2225 
2226 /*
2227  * NAME:	raid_done
2228  * DESCRIPTION: RAID metadevice I/O done interrupt routine
2229  * PARAMETERS:	struct buf *bp - pointer to a buffer structure
2230  */
2231 static void
2232 raid_done(struct buf *bp)
2233 {
2234 	md_raidcs_t	*cs;
2235 	int		flags, frags;
2236 
2237 	sema_v(&bp->b_io);
2238 	cs = (md_raidcs_t *)bp->b_chain;
2239 
2240 	ASSERT(cs != NULL);
2241 
2242 	mutex_enter(&cs->cs_mx);
2243 	if (bp->b_flags & B_ERROR) {
2244 		cs->cs_flags |= MD_RCS_ERROR;
2245 		cs->cs_flags &= ~(MD_RCS_ISCALL);
2246 	}
2247 
2248 	flags = cs->cs_flags;
2249 	frags = --cs->cs_frags;
2250 	mutex_exit(&cs->cs_mx);
2251 	if (frags != 0) {
2252 		return;
2253 	}
2254 
2255 	if (flags & MD_RCS_ERROR) {
2256 		if (cs->cs_error_call) {
2257 			daemon_request(&md_done_daemon, cs->cs_error_call,
2258 			    (daemon_queue_t *)cs, REQ_OLD);
2259 		}
2260 		return;
2261 	}
2262 
2263 	if (flags & MD_RCS_ISCALL) {
2264 		cs->cs_flags &= ~(MD_RCS_ISCALL);
2265 		(*(cs->cs_call))(cs);
2266 		return;
2267 	}
2268 	daemon_request(&md_done_daemon, cs->cs_call,
2269 	    (daemon_queue_t *)cs, REQ_OLD);
2270 }
2271 /*
2272  * the flag RIO_EXTRA is used when dealing with a column in the process
2273  * of being resynced. During the resync, writes may have to take place
2274  * on both the original component and a hotspare component.
2275  */
2276 #define	RIO_DATA	0x00100		/* use data buffer & data column */
2277 #define	RIO_PARITY	0x00200		/* use parity buffer & parity column */
2278 #define	RIO_WRITE	0x00400		/* issue a write */
2279 #define	RIO_READ	0x00800		/* issue a read */
2280 #define	RIO_PWIO	0x01000		/* do the I/O to the prewrite entry */
2281 #define	RIO_ALT		0x02000		/* do write to alternate device */
2282 #define	RIO_EXTRA	0x04000		/* use extra buffer */
2283 
2284 #define	RIO_COLMASK	0x000ff
2285 
2286 #define	RIO_PREWRITE	RIO_WRITE | RIO_PWIO
2287 
2288 /*
2289  * NAME:	raidio
2290  * DESCRIPTION: RAID metadevice write routine
2291  * PARAMETERS:	md_raidcs_t *cs - pointer to a child structure
2292  */
2293 static void
2294 raidio(md_raidcs_t *cs, int flags)
2295 {
2296 	buf_t		*bp;
2297 	int		column;
2298 	int		flag;
2299 	void		*private;
2300 	mr_unit_t	*un;
2301 	int		iosize;
2302 	diskaddr_t	pwstart;
2303 	diskaddr_t	devstart;
2304 	md_dev64_t	dev;
2305 
2306 	un = cs->cs_un;
2307 
2308 	ASSERT(IO_READER_HELD(un));
2309 	ASSERT(UNIT_READER_HELD(un));
2310 
2311 	if (flags & RIO_DATA) {
2312 		if (flags & RIO_EXTRA)
2313 			bp = &cs->cs_hbuf;
2314 		else
2315 			bp = &cs->cs_dbuf;
2316 		bp->b_un.b_addr = cs->cs_dbuffer;
2317 		column = cs->cs_dcolumn;
2318 	} else {
2319 		if (flags & RIO_EXTRA)
2320 			bp = &cs->cs_hbuf;
2321 		else
2322 			bp = &cs->cs_pbuf;
2323 		bp->b_un.b_addr = cs->cs_pbuffer;
2324 		column = cs->cs_pcolumn;
2325 	}
2326 	if (flags & RIO_COLMASK)
2327 		column = (flags & RIO_COLMASK) - 1;
2328 
2329 	bp->b_bcount = cs->cs_bcount;
2330 	bp->b_bufsize = cs->cs_bcount;
2331 	iosize = un->un_iosize;
2332 
2333 	/* check if the hotspared device will be used */
2334 	if (flags & RIO_ALT && (flags & RIO_WRITE)) {
2335 		pwstart = un->un_column[column].un_alt_pwstart;
2336 		devstart = un->un_column[column].un_alt_devstart;
2337 		dev = un->un_column[column].un_alt_dev;
2338 	} else {
2339 		pwstart = un->un_column[column].un_pwstart;
2340 		devstart = un->un_column[column].un_devstart;
2341 		dev = un->un_column[column].un_dev;
2342 	}
2343 
2344 	/* if not writing to log skip log header */
2345 	if ((flags & RIO_PWIO) == 0) {
2346 		bp->b_lblkno = devstart + cs->cs_blkno;
2347 		bp->b_un.b_addr += DEV_BSIZE;
2348 	} else {
2349 		bp->b_bcount += DEV_BSIZE;
2350 		bp->b_bufsize = bp->b_bcount;
2351 		if (flags & RIO_DATA) {
2352 			bp->b_lblkno = cs->cs_dpwslot * iosize + pwstart;
2353 		} else { /* not DATA -> PARITY */
2354 			bp->b_lblkno = cs->cs_ppwslot * iosize + pwstart;
2355 		}
2356 	}
2357 
2358 	bp->b_flags &= ~(B_READ | B_WRITE | B_ERROR | nv_available);
2359 	bp->b_flags |= B_BUSY;
2360 	if (flags & RIO_READ) {
2361 		bp->b_flags |= B_READ;
2362 	} else {
2363 		bp->b_flags |= B_WRITE;
2364 		if ((nv_available && nv_parity && (flags & RIO_PARITY)) ||
2365 		    (nv_available && nv_prewrite && (flags & RIO_PWIO)))
2366 			bp->b_flags |= nv_available;
2367 	}
2368 	bp->b_iodone = (int (*)())raid_done;
2369 	bp->b_edev = md_dev64_to_dev(dev);
2370 
2371 	ASSERT((bp->b_edev != 0) && (bp->b_edev != NODEV));
2372 
2373 	private = cs->cs_strategy_private;
2374 	flag = cs->cs_strategy_flag;
2375 
2376 	md_call_strategy(bp, flag, private);
2377 }
2378 
2379 /*
2380  * NAME:	genstandardparity
2381  * DESCRIPTION: This routine
2382  * PARAMETERS:	md_raidcs_t *cs - pointer to a child structure
2383  */
2384 static void
2385 genstandardparity(md_raidcs_t *cs)
2386 {
2387 	uint_t		*dbuf, *pbuf;
2388 	size_t		wordcnt;
2389 	uint_t		dsum = 0;
2390 	uint_t		psum = 0;
2391 
2392 	ASSERT((cs->cs_bcount & 0x3) == 0);
2393 
2394 	wordcnt = cs->cs_bcount / sizeof (uint_t);
2395 
2396 	dbuf = (uint_t *)(void *)(cs->cs_dbuffer + DEV_BSIZE);
2397 	pbuf = (uint_t *)(void *)(cs->cs_pbuffer + DEV_BSIZE);
2398 
2399 	/* Word aligned */
2400 	if (((uintptr_t)cs->cs_addr & 0x3) == 0) {
2401 		uint_t	*uwbuf = (uint_t *)(void *)(cs->cs_addr);
2402 		uint_t	uval;
2403 
2404 		while (wordcnt--) {
2405 			uval = *uwbuf++;
2406 			psum ^= (*pbuf = ((*pbuf ^ *dbuf) ^ uval));
2407 			++pbuf;
2408 			*dbuf = uval;
2409 			dsum ^= uval;
2410 			++dbuf;
2411 		}
2412 	} else {
2413 		uchar_t	*ubbuf = (uchar_t *)(cs->cs_addr);
2414 		union {
2415 			uint_t	wb;
2416 			uchar_t	bb[4];
2417 		} cb;
2418 
2419 		while (wordcnt--) {
2420 			cb.bb[0] = *ubbuf++;
2421 			cb.bb[1] = *ubbuf++;
2422 			cb.bb[2] = *ubbuf++;
2423 			cb.bb[3] = *ubbuf++;
2424 			psum ^= (*pbuf = ((*pbuf ^ *dbuf) ^ cb.wb));
2425 			++pbuf;
2426 			*dbuf = cb.wb;
2427 			dsum ^= cb.wb;
2428 			++dbuf;
2429 		}
2430 	}
2431 
2432 	RAID_FILLIN_RPW(cs->cs_dbuffer, cs->cs_un, dsum, cs->cs_pcolumn,
2433 	    cs->cs_blkno, cs->cs_blkcnt, cs->cs_pwid,
2434 	    2, cs->cs_dcolumn, RAID_PWMAGIC);
2435 
2436 	RAID_FILLIN_RPW(cs->cs_pbuffer, cs->cs_un, psum, cs->cs_dcolumn,
2437 	    cs->cs_blkno, cs->cs_blkcnt, cs->cs_pwid,
2438 	    2, cs->cs_pcolumn, RAID_PWMAGIC);
2439 }
2440 
2441 static void
2442 genlineparity(md_raidcs_t *cs)
2443 {
2444 
2445 	mr_unit_t	*un = cs->cs_un;
2446 	md_raidcbuf_t	*cbuf;
2447 	uint_t		*pbuf, *dbuf;
2448 	uint_t		*uwbuf;
2449 	uchar_t		*ubbuf;
2450 	size_t		wordcnt;
2451 	uint_t		psum = 0, dsum = 0;
2452 	size_t		count = un->un_segsize * DEV_BSIZE;
2453 	uint_t		col;
2454 	buf_t		*bp;
2455 
2456 	ASSERT((cs->cs_bcount & 0x3) == 0);
2457 
2458 	pbuf = (uint_t *)(void *)(cs->cs_pbuffer + DEV_BSIZE);
2459 	dbuf = (uint_t *)(void *)(cs->cs_dbuffer + DEV_BSIZE);
2460 	uwbuf = (uint_t *)(void *)(cs->cs_addr);
2461 	ubbuf = (uchar_t *)(void *)(cs->cs_addr);
2462 
2463 	wordcnt = count / sizeof (uint_t);
2464 
2465 	/* Word aligned */
2466 	if (((uintptr_t)cs->cs_addr & 0x3) == 0) {
2467 		uint_t	 uval;
2468 
2469 		while (wordcnt--) {
2470 			uval = *uwbuf++;
2471 			*dbuf = uval;
2472 			*pbuf = uval;
2473 			dsum ^= uval;
2474 			++pbuf;
2475 			++dbuf;
2476 		}
2477 	} else {
2478 		union {
2479 			uint_t	wb;
2480 			uchar_t	bb[4];
2481 		} cb;
2482 
2483 		while (wordcnt--) {
2484 			cb.bb[0] = *ubbuf++;
2485 			cb.bb[1] = *ubbuf++;
2486 			cb.bb[2] = *ubbuf++;
2487 			cb.bb[3] = *ubbuf++;
2488 			*dbuf = cb.wb;
2489 			*pbuf = cb.wb;
2490 			dsum ^= cb.wb;
2491 			++pbuf;
2492 			++dbuf;
2493 		}
2494 	}
2495 
2496 	RAID_FILLIN_RPW(cs->cs_dbuffer, un, dsum, cs->cs_pcolumn,
2497 	    cs->cs_blkno, cs->cs_blkcnt, cs->cs_pwid,
2498 	    un->un_totalcolumncnt, cs->cs_dcolumn, RAID_PWMAGIC);
2499 
2500 	raidio(cs, RIO_PREWRITE | RIO_DATA);
2501 
2502 	for (cbuf = cs->cs_buflist; cbuf; cbuf = cbuf->cbuf_next) {
2503 
2504 		dsum = 0;
2505 		pbuf = (uint_t *)(void *)(cs->cs_pbuffer + DEV_BSIZE);
2506 		dbuf = (uint_t *)(void *)(cbuf->cbuf_buffer + DEV_BSIZE);
2507 
2508 		wordcnt = count / sizeof (uint_t);
2509 
2510 		col = cbuf->cbuf_column;
2511 
2512 		/* Word aligned */
2513 		if (((uintptr_t)cs->cs_addr & 0x3) == 0) {
2514 			uint_t	uval;
2515 
2516 			/*
2517 			 * Only calculate psum when working on the last
2518 			 * data buffer.
2519 			 */
2520 			if (cbuf->cbuf_next == NULL) {
2521 				psum = 0;
2522 				while (wordcnt--) {
2523 					uval = *uwbuf++;
2524 					*dbuf = uval;
2525 					psum ^= (*pbuf ^= uval);
2526 					dsum ^= uval;
2527 					++dbuf;
2528 					++pbuf;
2529 				}
2530 			} else {
2531 				while (wordcnt--) {
2532 					uval = *uwbuf++;
2533 					*dbuf = uval;
2534 					*pbuf ^= uval;
2535 					dsum ^= uval;
2536 					++dbuf;
2537 					++pbuf;
2538 				}
2539 			}
2540 		} else {
2541 			union {
2542 				uint_t	wb;
2543 				uchar_t	bb[4];
2544 			} cb;
2545 
2546 			/*
2547 			 * Only calculate psum when working on the last
2548 			 * data buffer.
2549 			 */
2550 			if (cbuf->cbuf_next == NULL) {
2551 				psum = 0;
2552 				while (wordcnt--) {
2553 					cb.bb[0] = *ubbuf++;
2554 					cb.bb[1] = *ubbuf++;
2555 					cb.bb[2] = *ubbuf++;
2556 					cb.bb[3] = *ubbuf++;
2557 					*dbuf = cb.wb;
2558 					psum ^= (*pbuf ^= cb.wb);
2559 					dsum ^= cb.wb;
2560 					++dbuf;
2561 					++pbuf;
2562 				}
2563 			} else {
2564 				while (wordcnt--) {
2565 					cb.bb[0] = *ubbuf++;
2566 					cb.bb[1] = *ubbuf++;
2567 					cb.bb[2] = *ubbuf++;
2568 					cb.bb[3] = *ubbuf++;
2569 					*dbuf = cb.wb;
2570 					*pbuf ^= cb.wb;
2571 					dsum ^= cb.wb;
2572 					++dbuf;
2573 					++pbuf;
2574 				}
2575 			}
2576 		}
2577 		RAID_FILLIN_RPW(cbuf->cbuf_buffer, un, dsum, cs->cs_pcolumn,
2578 		    cs->cs_blkno, cs->cs_blkcnt, cs->cs_pwid,
2579 		    un->un_totalcolumncnt, col, RAID_PWMAGIC);
2580 
2581 		/*
2582 		 * fill in buffer for write to prewrite area
2583 		 */
2584 		bp = &cbuf->cbuf_bp;
2585 		bp->b_un.b_addr = cbuf->cbuf_buffer;
2586 		bp->b_bcount = cbuf->cbuf_bcount + DEV_BSIZE;
2587 		bp->b_bufsize = bp->b_bcount;
2588 		bp->b_lblkno = (cbuf->cbuf_pwslot * un->un_iosize) +
2589 		    un->un_column[col].un_pwstart;
2590 		bp->b_flags = B_WRITE | B_BUSY;
2591 		if (nv_available && nv_prewrite)
2592 			bp->b_flags |= nv_available;
2593 		bp->b_iodone = (int (*)())raid_done;
2594 		bp->b_edev = md_dev64_to_dev(un->un_column[col].un_dev);
2595 		bp->b_chain = (struct buf *)cs;
2596 		md_call_strategy(bp,
2597 		    cs->cs_strategy_flag, cs->cs_strategy_private);
2598 	}
2599 
2600 	RAID_FILLIN_RPW(cs->cs_pbuffer, un, psum, cs->cs_dcolumn,
2601 	    cs->cs_blkno, cs->cs_blkcnt, cs->cs_pwid,
2602 	    un->un_totalcolumncnt, cs->cs_pcolumn, RAID_PWMAGIC);
2603 
2604 	raidio(cs, RIO_PREWRITE | RIO_PARITY);
2605 }
2606 
2607 /*
2608  * NAME:	raid_readregenloop
2609  * DESCRIPTION: RAID metadevice write routine
2610  * PARAMETERS:	md_raidcs_t *cs - pointer to a child structure
2611  */
2612 static void
2613 raid_readregenloop(md_raidcs_t *cs)
2614 {
2615 	mr_unit_t	*un;
2616 	md_raidps_t	*ps;
2617 	uint_t		*dbuf;
2618 	uint_t		*pbuf;
2619 	size_t		wordcnt;
2620 
2621 	un = cs->cs_un;
2622 
2623 	/*
2624 	 * XOR the parity with data bytes, must skip the
2625 	 * pre-write entry header in all data/parity buffers
2626 	 */
2627 	wordcnt = cs->cs_bcount / sizeof (uint_t);
2628 	dbuf = (uint_t *)(void *)(cs->cs_dbuffer + DEV_BSIZE);
2629 	pbuf = (uint_t *)(void *)(cs->cs_pbuffer + DEV_BSIZE);
2630 	while (wordcnt--)
2631 		*dbuf++ ^= *pbuf++;
2632 
2633 	/* bump up the loop count */
2634 	cs->cs_loop++;
2635 
2636 	/* skip the errored component */
2637 	if (cs->cs_loop == cs->cs_dcolumn)
2638 		cs->cs_loop++;
2639 
2640 	if (cs->cs_loop != un->un_totalcolumncnt) {
2641 		cs->cs_frags = 1;
2642 		raidio(cs, RIO_PARITY | RIO_READ | (cs->cs_loop + 1));
2643 		return;
2644 	}
2645 	/* reaching the end sof loop */
2646 	ps = cs->cs_ps;
2647 	bcopy(cs->cs_dbuffer + DEV_BSIZE, cs->cs_addr, cs->cs_bcount);
2648 	raid_free_child(cs, 1);
2649 
2650 	/* decrement readfrags */
2651 	raid_free_parent(ps, RFP_DECR_READFRAGS | RFP_RLS_LOCK);
2652 }
2653 
2654 /*
2655  * NAME:	raid_read_io
2656  * DESCRIPTION: RAID metadevice read I/O routine
2657  * PARAMETERS:	mr_unit_t *un - pointer to a unit structure
2658  *		md_raidcs_t *cs - pointer to a child structure
2659  */
2660 static void
2661 raid_read_io(mr_unit_t *un, md_raidcs_t *cs)
2662 {
2663 	int	flag;
2664 	void	*private;
2665 	buf_t	*bp;
2666 	buf_t	*pb = cs->cs_ps->ps_bp;
2667 	mr_column_t	*column;
2668 
2669 	flag = cs->cs_strategy_flag;
2670 	private = cs->cs_strategy_private;
2671 	column = &un->un_column[cs->cs_dcolumn];
2672 
2673 	/*
2674 	 * The component to be read is good, simply set up bp structure
2675 	 * and call low level md routine doing the read.
2676 	 */
2677 
2678 	if (COLUMN_ISOKAY(un, cs->cs_dcolumn) ||
2679 	    (COLUMN_ISLASTERR(un, cs->cs_dcolumn) &&
2680 	    (cs->cs_flags & MD_RCS_RECOVERY) == 0)) {
2681 		dev_t ddi_dev; /* needed for bioclone, so not md_dev64_t */
2682 		ddi_dev = md_dev64_to_dev(column->un_dev);
2683 
2684 		bp = &cs->cs_dbuf;
2685 		bp = md_bioclone(pb, cs->cs_offset, cs->cs_bcount, ddi_dev,
2686 		    column->un_devstart + cs->cs_blkno,
2687 		    (int (*)())raid_done, bp, KM_NOSLEEP);
2688 
2689 		bp->b_chain = (buf_t *)cs;
2690 
2691 		cs->cs_frags = 1;
2692 		cs->cs_error_call = raid_read_error;
2693 		cs->cs_retry_call = raid_read_retry;
2694 		cs->cs_flags |= MD_RCS_ISCALL;
2695 		cs->cs_stage = RAID_READ_DONE;
2696 		cs->cs_call = raid_stage;
2697 
2698 		ASSERT(bp->b_edev != 0);
2699 
2700 		md_call_strategy(bp, flag, private);
2701 		return;
2702 	}
2703 
2704 	/*
2705 	 * The component to be read is bad, have to go through
2706 	 * raid specific method to read data from other members.
2707 	 */
2708 	cs->cs_loop = 0;
2709 	/*
2710 	 * NOTE: always get dbuffer before pbuffer
2711 	 *	 and get both buffers before pwslot
2712 	 *	 otherwise a deadlock could be introduced.
2713 	 */
2714 	raid_mapin_buf(cs);
2715 	getdbuffer(cs);
2716 	getpbuffer(cs);
2717 	if (cs->cs_loop == cs->cs_dcolumn)
2718 		cs->cs_loop++;
2719 
2720 	/* zero out data buffer for use as a data sink */
2721 	bzero(cs->cs_dbuffer + DEV_BSIZE, cs->cs_bcount);
2722 	cs->cs_stage = RAID_NONE;
2723 	cs->cs_call = raid_readregenloop;
2724 	cs->cs_error_call = raid_read_error;
2725 	cs->cs_retry_call = raid_read_no_retry;
2726 	cs->cs_frags = 1;
2727 
2728 	/* use parity buffer to read other columns */
2729 	raidio(cs, RIO_PARITY | RIO_READ | (cs->cs_loop + 1));
2730 }
2731 
2732 /*
2733  * NAME:	raid_read
2734  * DESCRIPTION: RAID metadevice write routine
2735  * PARAMETERS:	mr_unit_t *un - pointer to a unit structure
2736  *		md_raidcs_t *cs - pointer to a child structure
2737  */
2738 static int
2739 raid_read(mr_unit_t *un, md_raidcs_t *cs)
2740 {
2741 	int		error = 0;
2742 	md_raidps_t	*ps;
2743 	mdi_unit_t	*ui;
2744 	minor_t		mnum;
2745 
2746 	ASSERT(IO_READER_HELD(un));
2747 	ps = cs->cs_ps;
2748 	ui = ps->ps_ui;
2749 	raid_line_reader_lock(cs, 0);
2750 	un = (mr_unit_t *)md_unit_readerlock(ui);
2751 	ASSERT(UNIT_STATE(un) != RUS_INIT);
2752 	mnum = MD_SID(un);
2753 	cs->cs_un = un;
2754 
2755 	/* make sure the read doesn't go beyond the end of the column */
2756 	if (cs->cs_blkno + cs->cs_blkcnt >
2757 	    un->un_segsize * un->un_segsincolumn) {
2758 		error = ENXIO;
2759 	}
2760 	if (error)
2761 		goto rerror;
2762 
2763 	if (un->un_state & RUS_REGEN) {
2764 		raid_regen_parity(cs);
2765 		un = MD_UNIT(mnum);
2766 		cs->cs_un = un;
2767 	}
2768 
2769 	raid_read_io(un, cs);
2770 	return (0);
2771 
2772 rerror:
2773 	raid_error_parent(ps, error);
2774 	raid_free_child(cs, 1);
2775 	/* decrement readfrags */
2776 	raid_free_parent(ps, RFP_DECR_READFRAGS | RFP_RLS_LOCK);
2777 	return (0);
2778 }
2779 
2780 /*
2781  * NAME:	raid_write_err_retry
2782  * DESCRIPTION: RAID metadevice write retry routine
2783  *		write was for parity or data only;
2784  *		complete write with error, no recovery possible
2785  * PARAMETERS:	mr_unit_t *un - pointer to a unit structure
2786  *		md_raidcs_t *cs - pointer to a child structure
2787  */
2788 /*ARGSUSED*/
2789 static void
2790 raid_write_err_retry(mr_unit_t *un, md_raidcs_t *cs)
2791 {
2792 	md_raidps_t	*ps = cs->cs_ps;
2793 	int		flags = RFP_DECR_FRAGS | RFP_RLS_LOCK;
2794 
2795 	/* decrement pwfrags if needed, and frags */
2796 	if (!(cs->cs_flags & MD_RCS_PWDONE))
2797 		flags |= RFP_DECR_PWFRAGS;
2798 	raid_error_parent(ps, EIO);
2799 	raid_free_child(cs, 1);
2800 	raid_free_parent(ps, flags);
2801 }
2802 
2803 /*
2804  * NAME:	raid_write_err_retry
2805  * DESCRIPTION: RAID metadevice write retry routine
2806  *		 write is too far along to retry and parent
2807  *		 has already been signaled with iodone.
2808  * PARAMETERS:	mr_unit_t *un - pointer to a unit structure
2809  *		md_raidcs_t *cs - pointer to a child structure
2810  */
2811 /*ARGSUSED*/
2812 static void
2813 raid_write_no_retry(mr_unit_t *un, md_raidcs_t *cs)
2814 {
2815 	md_raidps_t	*ps = cs->cs_ps;
2816 	int		flags = RFP_DECR_FRAGS | RFP_RLS_LOCK;
2817 
2818 	/* decrement pwfrags if needed, and frags */
2819 	if (!(cs->cs_flags & MD_RCS_PWDONE))
2820 		flags |= RFP_DECR_PWFRAGS;
2821 	raid_free_child(cs, 1);
2822 	raid_free_parent(ps, flags);
2823 }
2824 
2825 /*
2826  * NAME:	raid_write_retry
2827  * DESCRIPTION: RAID metadevice write retry routine
2828  * PARAMETERS:	mr_unit_t *un - pointer to a unit structure
2829  *		md_raidcs_t *cs - pointer to a child structure
2830  */
2831 static void
2832 raid_write_retry(mr_unit_t *un, md_raidcs_t *cs)
2833 {
2834 	md_raidps_t	*ps;
2835 
2836 	ps = cs->cs_ps;
2837 
2838 	/* re-initialize the buf_t structure for raid_write() */
2839 	cs->cs_dbuf.b_chain = (struct buf *)cs;
2840 	cs->cs_dbuf.b_back = &cs->cs_dbuf;
2841 	cs->cs_dbuf.b_forw = &cs->cs_dbuf;
2842 	cs->cs_dbuf.b_flags = B_BUSY;	/* initialize flags */
2843 	cs->cs_dbuf.b_error = 0;	/* initialize error */
2844 	cs->cs_dbuf.b_offset = -1;
2845 	/* Initialize semaphores */
2846 	sema_init(&cs->cs_dbuf.b_io, 0, NULL,
2847 	    SEMA_DEFAULT, NULL);
2848 	sema_init(&cs->cs_dbuf.b_sem, 0, NULL,
2849 	    SEMA_DEFAULT, NULL);
2850 
2851 	cs->cs_pbuf.b_chain = (struct buf *)cs;
2852 	cs->cs_pbuf.b_back = &cs->cs_pbuf;
2853 	cs->cs_pbuf.b_forw = &cs->cs_pbuf;
2854 	cs->cs_pbuf.b_flags = B_BUSY;	/* initialize flags */
2855 	cs->cs_pbuf.b_error = 0;	/* initialize error */
2856 	cs->cs_pbuf.b_offset = -1;
2857 	sema_init(&cs->cs_pbuf.b_io, 0, NULL,
2858 	    SEMA_DEFAULT, NULL);
2859 	sema_init(&cs->cs_pbuf.b_sem, 0, NULL,
2860 	    SEMA_DEFAULT, NULL);
2861 
2862 	cs->cs_hbuf.b_chain = (struct buf *)cs;
2863 	cs->cs_hbuf.b_back = &cs->cs_hbuf;
2864 	cs->cs_hbuf.b_forw = &cs->cs_hbuf;
2865 	cs->cs_hbuf.b_flags = B_BUSY;	/* initialize flags */
2866 	cs->cs_hbuf.b_error = 0;	/* initialize error */
2867 	cs->cs_hbuf.b_offset = -1;
2868 	sema_init(&cs->cs_hbuf.b_io, 0, NULL,
2869 	    SEMA_DEFAULT, NULL);
2870 	sema_init(&cs->cs_hbuf.b_sem, 0, NULL,
2871 	    SEMA_DEFAULT, NULL);
2872 
2873 	cs->cs_flags &= ~(MD_RCS_ERROR);
2874 	/*
2875 	 * If we have already done'ed the i/o but have done prewrite
2876 	 * on this child, then reset PWDONE flag and bump pwfrags before
2877 	 * restarting i/o.
2878 	 * If pwfrags is zero, we have already 'iodone'd the i/o so
2879 	 * leave things alone.  We don't want to re-'done' it.
2880 	 */
2881 	mutex_enter(&ps->ps_mx);
2882 	if (cs->cs_flags & MD_RCS_PWDONE) {
2883 		cs->cs_flags &= ~MD_RCS_PWDONE;
2884 		ps->ps_pwfrags++;
2885 	}
2886 	mutex_exit(&ps->ps_mx);
2887 	raid_write_io(un, cs);
2888 }
2889 
2890 /*
2891  * NAME:	raid_wrerr
2892  * DESCRIPTION: RAID metadevice write routine
2893  * PARAMETERS:	md_raidcs_t *cs - pointer to a child structure
2894  * LOCKS:	must obtain unit writer lock while calling raid_error_state
2895  *		since a unit or column state transition may take place.
2896  *		must obtain unit reader lock to retry I/O.
2897  */
2898 static void
2899 raid_wrerr(md_raidcs_t *cs)
2900 {
2901 	md_raidps_t	*ps;
2902 	mdi_unit_t	*ui;
2903 	mr_unit_t	*un;
2904 	md_raidcbuf_t	*cbuf;
2905 
2906 	ps = cs->cs_ps;
2907 	ui = ps->ps_ui;
2908 
2909 	un = (mr_unit_t *)md_unit_writerlock(ui);
2910 	ASSERT(un != 0);
2911 
2912 	if (cs->cs_dbuf.b_flags & B_ERROR)
2913 		(void) raid_error_state(un, &cs->cs_dbuf);
2914 	if (cs->cs_pbuf.b_flags & B_ERROR)
2915 		(void) raid_error_state(un, &cs->cs_pbuf);
2916 	if (cs->cs_hbuf.b_flags & B_ERROR)
2917 		(void) raid_error_state(un, &cs->cs_hbuf);
2918 	for (cbuf = cs->cs_buflist; cbuf; cbuf = cbuf->cbuf_next)
2919 		if (cbuf->cbuf_bp.b_flags & B_ERROR)
2920 			(void) raid_error_state(un, &cbuf->cbuf_bp);
2921 
2922 	md_unit_writerexit(ui);
2923 
2924 	ps->ps_flags |= MD_RPS_HSREQ;
2925 
2926 	un = (mr_unit_t *)md_unit_readerlock(ui);
2927 
2928 	/* now attempt the appropriate retry routine */
2929 	(*(cs->cs_retry_call))(un, cs);
2930 }
2931 /*
2932  * NAMES:	raid_write_error
2933  * DESCRIPTION: I/O error handling routine for a RAID metadevice write
2934  * PARAMETERS:	md_raidcs_t *cs - pointer to child structure
2935  */
2936 /*ARGSUSED*/
2937 static void
2938 raid_write_error(md_raidcs_t *cs)
2939 {
2940 	md_raidps_t	*ps;
2941 	mdi_unit_t	*ui;
2942 	mr_unit_t	*un;
2943 	md_raidcbuf_t	*cbuf;
2944 	set_t		setno;
2945 
2946 	ps = cs->cs_ps;
2947 	ui = ps->ps_ui;
2948 	un = cs->cs_un;
2949 
2950 	setno = MD_UN2SET(un);
2951 
2952 	/*
2953 	 * locate each buf that is in error on this io and then
2954 	 * output an error message
2955 	 */
2956 	if ((cs->cs_dbuf.b_flags & B_ERROR) &&
2957 	    (COLUMN_STATE(un, cs->cs_dcolumn) != RCS_ERRED) &&
2958 	    (COLUMN_STATE(un, cs->cs_dcolumn) != RCS_LAST_ERRED))
2959 		cmn_err(CE_WARN, "md %s: write error on %s",
2960 		    md_shortname(MD_SID(un)),
2961 		    md_devname(setno, md_expldev(cs->cs_dbuf.b_edev), NULL, 0));
2962 
2963 	if ((cs->cs_pbuf.b_flags & B_ERROR) &&
2964 	    (COLUMN_STATE(un, cs->cs_pcolumn) != RCS_ERRED) &&
2965 	    (COLUMN_STATE(un, cs->cs_pcolumn) != RCS_LAST_ERRED))
2966 		cmn_err(CE_WARN, "md %s: write error on %s",
2967 		    md_shortname(MD_SID(un)),
2968 		    md_devname(setno, md_expldev(cs->cs_pbuf.b_edev), NULL, 0));
2969 
2970 	for (cbuf = cs->cs_buflist; cbuf; cbuf = cbuf->cbuf_next)
2971 		if ((cbuf->cbuf_bp.b_flags & B_ERROR) &&
2972 		    (COLUMN_STATE(un, cbuf->cbuf_column) != RCS_ERRED) &&
2973 		    (COLUMN_STATE(un, cbuf->cbuf_column) != RCS_LAST_ERRED))
2974 			cmn_err(CE_WARN, "md %s: write error on %s",
2975 			    md_shortname(MD_SID(un)),
2976 			    md_devname(setno, md_expldev(cbuf->cbuf_bp.b_edev),
2977 			    NULL, 0));
2978 
2979 	md_unit_readerexit(ui);
2980 
2981 	ASSERT(cs->cs_frags == 0);
2982 
2983 	/* now schedule processing for possible state change */
2984 	daemon_request(&md_mstr_daemon, raid_wrerr,
2985 	    (daemon_queue_t *)cs, REQ_OLD);
2986 
2987 }
2988 
2989 /*
2990  * NAME:	raid_write_ponly
2991  * DESCRIPTION: RAID metadevice write routine
2992  *		in the case where only the parity column can be written
2993  * PARAMETERS:	md_raidcs_t *cs - pointer to a child structure
2994  */
2995 static void
2996 raid_write_ponly(md_raidcs_t *cs)
2997 {
2998 	md_raidps_t	*ps;
2999 	mr_unit_t	*un = cs->cs_un;
3000 
3001 	ps = cs->cs_ps;
3002 	/* decrement pwfrags if needed, but not frags */
3003 	ASSERT(!(cs->cs_flags & MD_RCS_PWDONE));
3004 	raid_free_parent(ps, RFP_DECR_PWFRAGS);
3005 	cs->cs_flags |= MD_RCS_PWDONE;
3006 	cs->cs_frags = 1;
3007 	cs->cs_stage = RAID_WRITE_PONLY_DONE;
3008 	cs->cs_call = raid_stage;
3009 	cs->cs_error_call = raid_write_error;
3010 	cs->cs_retry_call = raid_write_no_retry;
3011 	if (WRITE_ALT(un, cs->cs_pcolumn)) {
3012 		cs->cs_frags++;
3013 		raidio(cs, RIO_ALT | RIO_EXTRA | RIO_PARITY | RIO_WRITE);
3014 	}
3015 	raidio(cs, RIO_PARITY | RIO_WRITE);
3016 }
3017 
3018 /*
3019  * NAME:	raid_write_ploop
3020  * DESCRIPTION: RAID metadevice write routine, constructs parity from
3021  *		data in other columns.
3022  * PARAMETERS:	md_raidcs_t *cs - pointer to a child structure
3023  */
3024 static void
3025 raid_write_ploop(md_raidcs_t *cs)
3026 {
3027 	mr_unit_t *un = cs->cs_un;
3028 	uint_t *dbuf;
3029 	uint_t *pbuf;
3030 	size_t wordcnt;
3031 	uint_t psum = 0;
3032 
3033 	wordcnt = cs->cs_bcount / sizeof (uint_t);
3034 	dbuf = (uint_t *)(void *)(cs->cs_dbuffer + DEV_BSIZE);
3035 	pbuf = (uint_t *)(void *)(cs->cs_pbuffer + DEV_BSIZE);
3036 	while (wordcnt--)
3037 		*pbuf++ ^= *dbuf++;
3038 	cs->cs_loop++;
3039 
3040 	/*
3041 	 * build parity from scratch using new data,
3042 	 * skip reading the data and parity columns.
3043 	 */
3044 	while (cs->cs_loop == cs->cs_dcolumn || cs->cs_loop == cs->cs_pcolumn)
3045 		cs->cs_loop++;
3046 
3047 	if (cs->cs_loop != un->un_totalcolumncnt) {
3048 		cs->cs_frags = 1;
3049 		raidio(cs, RIO_DATA | RIO_READ | (cs->cs_loop + 1));
3050 		return;
3051 	}
3052 
3053 	/* construct checksum for parity buffer */
3054 	wordcnt = cs->cs_bcount / sizeof (uint_t);
3055 	pbuf = (uint_t *)(void *)(cs->cs_pbuffer + DEV_BSIZE);
3056 	while (wordcnt--) {
3057 		psum ^= *pbuf;
3058 		pbuf++;
3059 	}
3060 	RAID_FILLIN_RPW(cs->cs_pbuffer, un, psum, -1,
3061 	    cs->cs_blkno, cs->cs_blkcnt, cs->cs_pwid,
3062 	    1, cs->cs_pcolumn, RAID_PWMAGIC);
3063 
3064 	cs->cs_stage = RAID_NONE;
3065 	cs->cs_call = raid_write_ponly;
3066 	cs->cs_error_call = raid_write_error;
3067 	cs->cs_retry_call = raid_write_err_retry;
3068 	cs->cs_frags = 1;
3069 	if (WRITE_ALT(un, cs->cs_pcolumn)) {
3070 		cs->cs_frags++;
3071 		raidio(cs, RIO_ALT | RIO_EXTRA | RIO_PARITY | RIO_PREWRITE);
3072 	}
3073 	raidio(cs, RIO_PARITY | RIO_PREWRITE);
3074 }
3075 
3076 /*
3077  * NAME:	raid_write_donly
3078  * DESCRIPTION: RAID metadevice write routine
3079  *		Completed writing data to prewrite entry
3080  *		in the case where only the data column can be written
3081  * PARAMETERS:	md_raidcs_t *cs - pointer to a child structure
3082  */
3083 static void
3084 raid_write_donly(md_raidcs_t *cs)
3085 {
3086 	md_raidps_t	*ps;
3087 	mr_unit_t	*un = cs->cs_un;
3088 
3089 	ps = cs->cs_ps;
3090 	/* WARNING: don't release unit reader lock here... */
3091 	/* decrement pwfrags if needed, but not frags */
3092 	ASSERT(!(cs->cs_flags & MD_RCS_PWDONE));
3093 	raid_free_parent(ps, RFP_DECR_PWFRAGS);
3094 	cs->cs_flags |= MD_RCS_PWDONE;
3095 	cs->cs_frags = 1;
3096 	cs->cs_stage = RAID_WRITE_DONLY_DONE;
3097 	cs->cs_call = raid_stage;
3098 	cs->cs_error_call = raid_write_error;
3099 	cs->cs_retry_call = raid_write_err_retry;
3100 	if (WRITE_ALT(un, cs->cs_dcolumn)) {
3101 		cs->cs_frags++;
3102 		raidio(cs, RIO_ALT | RIO_EXTRA | RIO_DATA | RIO_WRITE);
3103 	}
3104 	raidio(cs, RIO_DATA | RIO_WRITE);
3105 }
3106 
3107 /*
3108  * NAME:	raid_write_got_old
3109  * DESCRIPTION: RAID metadevice write routine
3110  *		completed read of old data and old parity
3111  * PARAMETERS:	md_raidcs_t *cs - pointer to a child structure
3112  */
3113 static void
3114 raid_write_got_old(md_raidcs_t *cs)
3115 {
3116 	mr_unit_t *un = cs->cs_un;
3117 
3118 	ASSERT(IO_READER_HELD(cs->cs_un));
3119 	ASSERT(UNIT_READER_HELD(cs->cs_un));
3120 
3121 	raid_mapin_buf(cs);
3122 	genstandardparity(cs);
3123 	cs->cs_frags = 2;
3124 	cs->cs_call = raid_stage;
3125 	cs->cs_stage = RAID_PREWRITE_DONE;
3126 	cs->cs_error_call = raid_write_error;
3127 	cs->cs_retry_call = raid_write_retry;
3128 
3129 	if (WRITE_ALT(un, cs->cs_dcolumn)) {
3130 		cs->cs_frags++;
3131 		raidio(cs, RIO_ALT | RIO_EXTRA | RIO_DATA | RIO_PREWRITE);
3132 	}
3133 
3134 	if (WRITE_ALT(un, cs->cs_pcolumn)) {
3135 		cs->cs_frags++;
3136 		raidio(cs, RIO_ALT | RIO_EXTRA | RIO_PARITY | RIO_PREWRITE);
3137 	}
3138 	ASSERT(cs->cs_frags < 4);
3139 	raidio(cs,  RIO_DATA | RIO_PREWRITE);
3140 	raidio(cs,  RIO_PARITY | RIO_PREWRITE);
3141 }
3142 
3143 /*
3144  * NAME:	raid_write_io
3145  * DESCRIPTION: RAID metadevice write I/O routine
3146  * PARAMETERS:	mr_unit_t *un -  pointer to a unit structure
3147  *		md_raidcs_t *cs - pointer to a child structure
3148  */
3149 
3150 /*ARGSUSED*/
3151 static void
3152 raid_write_io(mr_unit_t *un, md_raidcs_t *cs)
3153 {
3154 	md_raidps_t	*ps = cs->cs_ps;
3155 	uint_t		*dbuf;
3156 	uint_t		*ubuf;
3157 	size_t		wordcnt;
3158 	uint_t		dsum = 0;
3159 	int		pcheck;
3160 	int		dcheck;
3161 
3162 	ASSERT((un->un_column[cs->cs_pcolumn].un_devstate &
3163 	    RCS_INIT) == 0);
3164 	ASSERT((un->un_column[cs->cs_dcolumn].un_devstate &
3165 	    RCS_INIT) == 0);
3166 	ASSERT(IO_READER_HELD(un));
3167 	ASSERT(UNIT_READER_HELD(un));
3168 	ASSERT(cs->cs_flags & MD_RCS_HAVE_PW_SLOTS);
3169 	if (cs->cs_flags & MD_RCS_LINE) {
3170 
3171 		mr_unit_t	*un = cs->cs_un;
3172 
3173 		ASSERT(un->un_origcolumncnt == un->un_totalcolumncnt);
3174 		raid_mapin_buf(cs);
3175 		cs->cs_frags = un->un_origcolumncnt;
3176 		cs->cs_call = raid_stage;
3177 		cs->cs_error_call = raid_write_error;
3178 		cs->cs_retry_call = raid_write_no_retry;
3179 		cs->cs_stage = RAID_LINE_PWDONE;
3180 		genlineparity(cs);
3181 		return;
3182 	}
3183 
3184 	pcheck = erred_check_line(un, cs, &un->un_column[cs->cs_pcolumn]);
3185 	dcheck = erred_check_line(un, cs, &un->un_column[cs->cs_dcolumn]);
3186 	cs->cs_resync_check = pcheck << RCL_PARITY_OFFSET || dcheck;
3187 
3188 	if (pcheck == RCL_ERRED && dcheck == RCL_ERRED) {
3189 		int err = EIO;
3190 
3191 		if ((un->un_column[cs->cs_pcolumn].un_devstate ==
3192 		    RCS_LAST_ERRED) ||
3193 		    (un->un_column[cs->cs_dcolumn].un_devstate ==
3194 		    RCS_LAST_ERRED))
3195 			err = ENXIO;
3196 		raid_error_parent(ps, err);
3197 		ASSERT(!(cs->cs_flags & MD_RCS_PWDONE));
3198 		raid_free_child(cs, 1);
3199 		raid_free_parent(ps,  RFP_DECR_FRAGS
3200 		    | RFP_RLS_LOCK | RFP_DECR_PWFRAGS);
3201 		return;
3202 	}
3203 
3204 	if (pcheck & RCL_ERRED) {
3205 		/*
3206 		 * handle case of only having data drive
3207 		 */
3208 		raid_mapin_buf(cs);
3209 		wordcnt = cs->cs_bcount / sizeof (uint_t);
3210 
3211 		dbuf = (uint_t *)(void *)(cs->cs_dbuffer + DEV_BSIZE);
3212 		ubuf = (uint_t *)(void *)(cs->cs_addr);
3213 
3214 		while (wordcnt--) {
3215 			*dbuf = *ubuf;
3216 			dsum ^= *ubuf;
3217 			dbuf++;
3218 			ubuf++;
3219 		}
3220 		RAID_FILLIN_RPW(cs->cs_dbuffer, un, dsum, -1,
3221 		    cs->cs_blkno, cs->cs_blkcnt, cs->cs_pwid,
3222 		    1, cs->cs_dcolumn, RAID_PWMAGIC);
3223 		cs->cs_frags = 1;
3224 		cs->cs_stage = RAID_NONE;
3225 		cs->cs_call = raid_write_donly;
3226 		cs->cs_error_call = raid_write_error;
3227 		cs->cs_retry_call = raid_write_err_retry;
3228 		if (WRITE_ALT(un, cs->cs_dcolumn)) {
3229 			cs->cs_frags++;
3230 			raidio(cs, RIO_DATA | RIO_ALT | RIO_EXTRA |
3231 			    RIO_PREWRITE);
3232 		}
3233 		raidio(cs, RIO_DATA | RIO_PREWRITE);
3234 		return;
3235 	}
3236 
3237 	if (dcheck & RCL_ERRED) {
3238 		/*
3239 		 * handle case of only having parity drive
3240 		 * build parity from scratch using new data,
3241 		 * skip reading the data and parity columns.
3242 		 */
3243 		raid_mapin_buf(cs);
3244 		cs->cs_loop = 0;
3245 		while (cs->cs_loop == cs->cs_dcolumn ||
3246 		    cs->cs_loop == cs->cs_pcolumn)
3247 			cs->cs_loop++;
3248 
3249 		/* copy new data in to begin building parity */
3250 		bcopy(cs->cs_addr, cs->cs_pbuffer + DEV_BSIZE, cs->cs_bcount);
3251 		cs->cs_stage = RAID_NONE;
3252 		cs->cs_call = raid_write_ploop;
3253 		cs->cs_error_call = raid_write_error;
3254 		cs->cs_retry_call = raid_write_err_retry;
3255 		cs->cs_frags = 1;
3256 		raidio(cs, RIO_DATA | RIO_READ | (cs->cs_loop + 1));
3257 		return;
3258 	}
3259 	/*
3260 	 * handle normal cases
3261 	 * read old data and old parity
3262 	 */
3263 	cs->cs_frags = 2;
3264 	cs->cs_stage = RAID_NONE;
3265 	cs->cs_call = raid_write_got_old;
3266 	cs->cs_error_call = raid_write_error;
3267 	cs->cs_retry_call = raid_write_retry;
3268 	ASSERT(ps->ps_magic == RAID_PSMAGIC);
3269 	raidio(cs, RIO_DATA | RIO_READ);
3270 	raidio(cs, RIO_PARITY | RIO_READ);
3271 }
3272 
3273 static void
3274 raid_enqueue(md_raidcs_t *cs)
3275 {
3276 	mdi_unit_t	*ui = cs->cs_ps->ps_ui;
3277 	kmutex_t	*io_list_mutex = &ui->ui_io_lock->io_list_mutex;
3278 	md_raidcs_t	*cs1;
3279 
3280 	mutex_enter(io_list_mutex);
3281 	ASSERT(! (cs->cs_flags & MD_RCS_LLOCKD));
3282 	if (ui->ui_io_lock->io_list_front == NULL) {
3283 		ui->ui_io_lock->io_list_front = cs;
3284 		ui->ui_io_lock->io_list_back = cs;
3285 	} else {
3286 		cs1 = ui->ui_io_lock->io_list_back;
3287 		cs1->cs_linlck_next = cs;
3288 		ui->ui_io_lock->io_list_back = cs;
3289 	}
3290 	STAT_INC(raid_write_waits);
3291 	STAT_MAX(raid_max_write_q_length, raid_write_queue_length);
3292 	cs->cs_linlck_next = NULL;
3293 	mutex_exit(io_list_mutex);
3294 }
3295 
3296 /*
3297  * NAME:	raid_write
3298  * DESCRIPTION: RAID metadevice write routine
3299  * PARAMETERS:	mr_unit_t *un -  pointer to a unit structure
3300  *		md_raidcs_t *cs - pointer to a child structure
3301  */
3302 
3303 /*ARGSUSED*/
3304 static int
3305 raid_write(mr_unit_t *un, md_raidcs_t *cs)
3306 {
3307 	int		error = 0;
3308 	md_raidps_t	*ps;
3309 	mdi_unit_t	*ui;
3310 	minor_t		mnum;
3311 
3312 	ASSERT(IO_READER_HELD(un));
3313 	ps = cs->cs_ps;
3314 	ui = ps->ps_ui;
3315 
3316 	ASSERT(UNIT_STATE(un) != RUS_INIT);
3317 	if (UNIT_STATE(un) == RUS_LAST_ERRED)
3318 		error = EIO;
3319 
3320 	/* make sure the write doesn't go beyond the column */
3321 	if (cs->cs_blkno + cs->cs_blkcnt > un->un_segsize * un->un_segsincolumn)
3322 		error = ENXIO;
3323 	if (error)
3324 		goto werror;
3325 
3326 	getresources(cs);
3327 
3328 	/*
3329 	 * this is an advisory loop that keeps the waiting lists short
3330 	 * to reduce cpu time.  Since there is a race introduced by not
3331 	 * aquiring all the correct mutexes, use a cv_timedwait to be
3332 	 * sure the write always will wake up and start.
3333 	 */
3334 	while (raid_check_pw(cs)) {
3335 		mutex_enter(&un->un_mx);
3336 		un->un_rflags |= MD_RFLAG_NEEDPW;
3337 		STAT_INC(raid_prewrite_waits);
3338 		(void) cv_reltimedwait(&un->un_cv, &un->un_mx, md_wr_wait,
3339 		    TR_CLOCK_TICK);
3340 		un->un_rflags &= ~MD_RFLAG_NEEDPW;
3341 		mutex_exit(&un->un_mx);
3342 	}
3343 
3344 	if (raid_line_writer_lock(cs, 1))
3345 		return (0);
3346 
3347 	un = (mr_unit_t *)md_unit_readerlock(ui);
3348 	cs->cs_un = un;
3349 	mnum = MD_SID(un);
3350 
3351 	if (un->un_state & RUS_REGEN) {
3352 		raid_regen_parity(cs);
3353 		un = MD_UNIT(mnum);
3354 		cs->cs_un = un;
3355 	}
3356 
3357 	raid_write_io(un, cs);
3358 	return (0);
3359 werror:
3360 	/* aquire unit reader lock sinc raid_free_child always drops it */
3361 	raid_error_parent(ps, error);
3362 	raid_free_child(cs, 0);
3363 	/* decrement both pwfrags and frags */
3364 	raid_free_parent(ps, RFP_DECR_PWFRAGS | RFP_DECR_FRAGS | RFP_RLS_LOCK);
3365 	return (0);
3366 }
3367 
3368 
3369 /*
3370  * NAMES:	raid_stage
3371  * DESCRIPTION: post-processing routine for a RAID metadevice
3372  * PARAMETERS:	md_raidcs_t *cs - pointer to child structure
3373  */
3374 static void
3375 raid_stage(md_raidcs_t *cs)
3376 {
3377 	md_raidps_t	*ps = cs->cs_ps;
3378 	mr_unit_t	*un = cs->cs_un;
3379 	md_raidcbuf_t	*cbuf;
3380 	buf_t		*bp;
3381 	void		*private;
3382 	int		flag;
3383 
3384 	switch (cs->cs_stage) {
3385 	case RAID_READ_DONE:
3386 		raid_free_child(cs, 1);
3387 		/* decrement readfrags */
3388 		raid_free_parent(ps, RFP_DECR_READFRAGS | RFP_RLS_LOCK);
3389 		return;
3390 
3391 	case RAID_WRITE_DONE:
3392 	case RAID_WRITE_PONLY_DONE:
3393 	case RAID_WRITE_DONLY_DONE:
3394 		/*
3395 		 *  Completed writing real parity and/or data.
3396 		 */
3397 		ASSERT(cs->cs_flags & MD_RCS_PWDONE);
3398 		raid_free_child(cs, 1);
3399 		/* decrement frags but not pwfrags */
3400 		raid_free_parent(ps, RFP_DECR_FRAGS | RFP_RLS_LOCK);
3401 		return;
3402 
3403 	case RAID_PREWRITE_DONE:
3404 		/*
3405 		 * completed writing data and parity to prewrite entries
3406 		 */
3407 		/*
3408 		 * WARNING: don't release unit reader lock here..
3409 		 * decrement pwfrags but not frags
3410 		 */
3411 		raid_free_parent(ps, RFP_DECR_PWFRAGS);
3412 		cs->cs_flags |= MD_RCS_PWDONE;
3413 		cs->cs_frags = 2;
3414 		cs->cs_stage = RAID_WRITE_DONE;
3415 		cs->cs_call = raid_stage;
3416 		cs->cs_error_call = raid_write_error;
3417 		cs->cs_retry_call = raid_write_no_retry;
3418 		if (WRITE_ALT(un, cs->cs_pcolumn)) {
3419 			cs->cs_frags++;
3420 			raidio(cs, RIO_ALT | RIO_EXTRA | RIO_PARITY |
3421 			    RIO_WRITE);
3422 		}
3423 		if (WRITE_ALT(un, cs->cs_dcolumn)) {
3424 			cs->cs_frags++;
3425 			raidio(cs, RIO_ALT | RIO_EXTRA | RIO_DATA | RIO_WRITE);
3426 		}
3427 		ASSERT(cs->cs_frags < 4);
3428 		raidio(cs, RIO_DATA | RIO_WRITE);
3429 		raidio(cs, RIO_PARITY | RIO_WRITE);
3430 		if (cs->cs_pw_inval_list) {
3431 			raid_free_pwinvalidate(cs);
3432 		}
3433 		return;
3434 
3435 	case RAID_LINE_PWDONE:
3436 		ASSERT(cs->cs_frags == 0);
3437 		raid_free_parent(ps, RFP_DECR_PWFRAGS);
3438 		cs->cs_flags |= MD_RCS_PWDONE;
3439 		cs->cs_frags = un->un_origcolumncnt;
3440 		cs->cs_call = raid_stage;
3441 		cs->cs_error_call = raid_write_error;
3442 		cs->cs_retry_call = raid_write_no_retry;
3443 		cs->cs_stage = RAID_WRITE_DONE;
3444 		for (cbuf = cs->cs_buflist; cbuf; cbuf = cbuf->cbuf_next) {
3445 			/*
3446 			 * fill in buffer for write to prewrite area
3447 			 */
3448 			bp = &cbuf->cbuf_bp;
3449 			bp->b_back = bp;
3450 			bp->b_forw = bp;
3451 			bp->b_un.b_addr = cbuf->cbuf_buffer + DEV_BSIZE;
3452 			bp->b_bcount = cbuf->cbuf_bcount;
3453 			bp->b_bufsize = cbuf->cbuf_bcount;
3454 			bp->b_lblkno =
3455 			    un->un_column[cbuf->cbuf_column].un_devstart +
3456 			    cs->cs_blkno;
3457 			bp->b_flags &= ~(B_READ | B_WRITE | B_ERROR);
3458 			bp->b_flags &= ~nv_available;
3459 			bp->b_flags |= B_WRITE | B_BUSY;
3460 			bp->b_iodone = (int (*)())raid_done;
3461 			bp->b_edev = md_dev64_to_dev(
3462 			    un->un_column[cbuf->cbuf_column].un_dev);
3463 			bp->b_chain = (struct buf *)cs;
3464 			private = cs->cs_strategy_private;
3465 			flag = cs->cs_strategy_flag;
3466 			md_call_strategy(bp, flag, private);
3467 		}
3468 		raidio(cs, RIO_DATA | RIO_WRITE);
3469 		raidio(cs, RIO_PARITY | RIO_WRITE);
3470 		if (cs->cs_pw_inval_list) {
3471 			raid_free_pwinvalidate(cs);
3472 		}
3473 		return;
3474 
3475 	default:
3476 		ASSERT(0);
3477 		break;
3478 	}
3479 }
3480 /*
3481  * NAME:	md_raid_strategy
3482  * DESCRIPTION: RAID metadevice I/O oprations entry point.
3483  * PARAMETERS:	buf_t	  *pb - pointer to a user I/O buffer
3484  *		int	 flag - metadevice specific flag
3485  *		void *private - carry over flag ??
3486  *
3487  */
3488 
3489 void
3490 md_raid_strategy(buf_t *pb, int flag, void *private)
3491 {
3492 	md_raidps_t	*ps;
3493 	md_raidcs_t	*cs;
3494 	int		doing_writes;
3495 	int		err;
3496 	mr_unit_t	*un;
3497 	mdi_unit_t	*ui;
3498 	size_t		count;
3499 	diskaddr_t	blkno;
3500 	caddr_t		addr;
3501 	off_t		offset;
3502 	int		colcnt;
3503 	minor_t		mnum;
3504 	set_t		setno;
3505 
3506 	ui = MDI_UNIT(getminor(pb->b_edev));
3507 	md_kstat_waitq_enter(ui);
3508 	un = (mr_unit_t *)md_io_readerlock(ui);
3509 	setno = MD_MIN2SET(getminor(pb->b_edev));
3510 
3511 	if ((flag & MD_NOBLOCK) == 0) {
3512 		if (md_inc_iocount(setno) != 0) {
3513 			pb->b_flags |= B_ERROR;
3514 			pb->b_error = ENXIO;
3515 			pb->b_resid = pb->b_bcount;
3516 			md_kstat_waitq_exit(ui);
3517 			md_io_readerexit(ui);
3518 			biodone(pb);
3519 			return;
3520 		}
3521 	} else {
3522 		md_inc_iocount_noblock(setno);
3523 	}
3524 
3525 	mnum = MD_SID(un);
3526 	colcnt = un->un_totalcolumncnt - 1;
3527 	count = pb->b_bcount;
3528 
3529 	STAT_CHECK(raid_512, count == 512);
3530 	STAT_CHECK(raid_1024, count == 1024);
3531 	STAT_CHECK(raid_1024_8192, count > 1024 && count < 8192);
3532 	STAT_CHECK(raid_8192, count == 8192);
3533 	STAT_CHECK(raid_8192_bigger, count > 8192);
3534 
3535 	(void *) md_unit_readerlock(ui);
3536 	if (!(flag & MD_STR_NOTTOP)) {
3537 		err = md_checkbuf(ui, (md_unit_t *)un, pb); /* check and map */
3538 		if (err != 0) {
3539 			md_kstat_waitq_exit(ui);
3540 			md_io_readerexit(ui);
3541 			return;
3542 		}
3543 	}
3544 	md_unit_readerexit(ui);
3545 
3546 	STAT_INC(raid_total_io);
3547 
3548 	/* allocate a parent structure for the user I/O */
3549 	ps = kmem_cache_alloc(raid_parent_cache, MD_ALLOCFLAGS);
3550 	raid_parent_init(ps);
3551 
3552 	/*
3553 	 * Save essential information from the original buffhdr
3554 	 * in the md_save structure.
3555 	 */
3556 	ps->ps_un = un;
3557 	ps->ps_ui = ui;
3558 	ps->ps_bp = pb;
3559 	ps->ps_addr = pb->b_un.b_addr;
3560 
3561 	if ((pb->b_flags & B_READ) == 0) {
3562 		ps->ps_flags |= MD_RPS_WRITE;
3563 		doing_writes = 1;
3564 		STAT_INC(raid_writes);
3565 	} else {
3566 		ps->ps_flags |= MD_RPS_READ;
3567 		doing_writes = 0;
3568 		STAT_INC(raid_reads);
3569 	}
3570 
3571 	count = lbtodb(pb->b_bcount);	/* transfer count (in blocks) */
3572 	blkno = pb->b_lblkno;		/* block number on device */
3573 	addr  = 0;
3574 	offset = 0;
3575 	ps->ps_pwfrags = 1;
3576 	ps->ps_frags = 1;
3577 	md_kstat_waitq_to_runq(ui);
3578 
3579 	do {
3580 		cs = kmem_cache_alloc(raid_child_cache, MD_ALLOCFLAGS);
3581 		raid_child_init(cs);
3582 		cs->cs_ps = ps;
3583 		cs->cs_un = un;
3584 		cs->cs_mdunit = mnum;
3585 		cs->cs_strategy_flag = flag;
3586 		cs->cs_strategy_private = private;
3587 		cs->cs_addr = addr;
3588 		cs->cs_offset = offset;
3589 		count = raid_iosetup(un, blkno, count, cs);
3590 		if (cs->cs_flags & MD_RCS_LINE) {
3591 			blkno += (cs->cs_blkcnt * colcnt);
3592 			offset += (cs->cs_bcount * colcnt);
3593 		} else {
3594 			blkno +=  cs->cs_blkcnt;
3595 			offset += cs->cs_bcount;
3596 		}
3597 		/* for each cs bump up the ps_pwfrags and ps_frags fields */
3598 		if (count) {
3599 			mutex_enter(&ps->ps_mx);
3600 			ps->ps_pwfrags++;
3601 			ps->ps_frags++;
3602 			mutex_exit(&ps->ps_mx);
3603 			if (doing_writes)
3604 				(void) raid_write(un, cs);
3605 			else
3606 				(void) raid_read(un, cs);
3607 		}
3608 	} while (count);
3609 	if (doing_writes) {
3610 		(void) raid_write(un, cs);
3611 	} else
3612 		(void) raid_read(un, cs);
3613 
3614 	if (! (flag & MD_STR_NOTTOP) && panicstr) {
3615 		while (! (ps->ps_flags & MD_RPS_DONE)) {
3616 			md_daemon(1, &md_done_daemon);
3617 			drv_usecwait(10);
3618 		}
3619 		kmem_cache_free(raid_parent_cache, ps);
3620 	}
3621 }
3622 
3623 /*
3624  * NAMES:	raid_snarf
3625  * DESCRIPTION: RAID metadevice SNARF entry point
3626  * PARAMETERS:	md_snarfcmd_t cmd,
3627  *		set_t setno
3628  * RETURNS:
3629  */
3630 static int
3631 raid_snarf(md_snarfcmd_t cmd, set_t setno)
3632 {
3633 	mr_unit_t	*un;
3634 	mddb_recid_t	recid;
3635 	int		gotsomething;
3636 	int		all_raid_gotten;
3637 	mddb_type_t	typ1;
3638 	uint_t		ncol;
3639 	mddb_de_ic_t	*dep;
3640 	mddb_rb32_t	*rbp;
3641 	size_t		newreqsize;
3642 	mr_unit_t	*big_un;
3643 	mr_unit32_od_t	*small_un;
3644 
3645 
3646 	if (cmd == MD_SNARF_CLEANUP)
3647 		return (0);
3648 
3649 	all_raid_gotten = 1;
3650 	gotsomething = 0;
3651 	typ1 = (mddb_type_t)md_getshared_key(setno,
3652 	    raid_md_ops.md_driver.md_drivername);
3653 	recid = mddb_makerecid(setno, 0);
3654 
3655 	while ((recid = mddb_getnextrec(recid, typ1, 0)) > 0) {
3656 		if (mddb_getrecprivate(recid) & MD_PRV_GOTIT) {
3657 			continue;
3658 		}
3659 
3660 		dep = mddb_getrecdep(recid);
3661 		dep->de_flags = MDDB_F_RAID;
3662 		rbp = dep->de_rb;
3663 		switch (rbp->rb_revision) {
3664 		case MDDB_REV_RB:
3665 		case MDDB_REV_RBFN:
3666 			if ((rbp->rb_private & MD_PRV_CONVD) == 0) {
3667 				/*
3668 				 * This means, we have an old and small record
3669 				 * and this record hasn't already been
3670 				 * converted.  Before we create an incore
3671 				 * metadevice from this we have to convert it to
3672 				 * a big record.
3673 				 */
3674 				small_un =
3675 				    (mr_unit32_od_t *)mddb_getrecaddr(recid);
3676 				ncol = small_un->un_totalcolumncnt;
3677 				newreqsize = sizeof (mr_unit_t) +
3678 				    ((ncol - 1) * sizeof (mr_column_t));
3679 				big_un = (mr_unit_t *)kmem_zalloc(newreqsize,
3680 				    KM_SLEEP);
3681 				raid_convert((caddr_t)small_un, (caddr_t)big_un,
3682 				    SMALL_2_BIG);
3683 				kmem_free(small_un, dep->de_reqsize);
3684 				dep->de_rb_userdata = big_un;
3685 				dep->de_reqsize = newreqsize;
3686 				un = big_un;
3687 				rbp->rb_private |= MD_PRV_CONVD;
3688 			} else {
3689 				/*
3690 				 * Record has already been converted.  Just
3691 				 * get its address.
3692 				 */
3693 				un = (mr_unit_t *)mddb_getrecaddr(recid);
3694 			}
3695 			un->c.un_revision &= ~MD_64BIT_META_DEV;
3696 			break;
3697 		case MDDB_REV_RB64:
3698 		case MDDB_REV_RB64FN:
3699 			/* Big device */
3700 			un = (mr_unit_t *)mddb_getrecaddr(recid);
3701 			un->c.un_revision |= MD_64BIT_META_DEV;
3702 			un->c.un_flag |= MD_EFILABEL;
3703 			break;
3704 		}
3705 		MDDB_NOTE_FN(rbp->rb_revision, un->c.un_revision);
3706 
3707 		/*
3708 		 * Create minor device node for snarfed entry.
3709 		 */
3710 		(void) md_create_minor_node(MD_MIN2SET(MD_SID(un)), MD_SID(un));
3711 
3712 		if (MD_UNIT(MD_SID(un)) != NULL) {
3713 			mddb_setrecprivate(recid, MD_PRV_PENDDEL);
3714 			continue;
3715 		}
3716 		all_raid_gotten = 0;
3717 		if (raid_build_incore((void *)un, 1) == 0) {
3718 			mddb_setrecprivate(recid, MD_PRV_GOTIT);
3719 			md_create_unit_incore(MD_SID(un), &raid_md_ops, 1);
3720 			gotsomething = 1;
3721 		} else if (un->mr_ic) {
3722 			kmem_free(un->un_column_ic, sizeof (mr_column_ic_t) *
3723 			    un->un_totalcolumncnt);
3724 			kmem_free(un->mr_ic, sizeof (*un->mr_ic));
3725 		}
3726 	}
3727 
3728 	if (!all_raid_gotten) {
3729 		return (gotsomething);
3730 	}
3731 
3732 	recid = mddb_makerecid(setno, 0);
3733 	while ((recid = mddb_getnextrec(recid, typ1, 0)) > 0)
3734 		if (!(mddb_getrecprivate(recid) & MD_PRV_GOTIT))
3735 			mddb_setrecprivate(recid, MD_PRV_PENDDEL);
3736 
3737 	return (0);
3738 }
3739 
3740 /*
3741  * NAMES:	raid_halt
3742  * DESCRIPTION: RAID metadevice HALT entry point
3743  * PARAMETERS:	md_haltcmd_t cmd -
3744  *		set_t	setno -
3745  * RETURNS:
3746  */
3747 static int
3748 raid_halt(md_haltcmd_t cmd, set_t setno)
3749 {
3750 	set_t		i;
3751 	mdi_unit_t	*ui;
3752 	minor_t		mnum;
3753 
3754 	if (cmd == MD_HALT_CLOSE)
3755 		return (0);
3756 
3757 	if (cmd == MD_HALT_OPEN)
3758 		return (0);
3759 
3760 	if (cmd == MD_HALT_UNLOAD)
3761 		return (0);
3762 
3763 	if (cmd == MD_HALT_CHECK) {
3764 		for (i = 0; i < md_nunits; i++) {
3765 			mnum = MD_MKMIN(setno, i);
3766 			if ((ui = MDI_UNIT(mnum)) == NULL)
3767 				continue;
3768 			if (ui->ui_opsindex != raid_md_ops.md_selfindex)
3769 				continue;
3770 			if (md_unit_isopen(ui))
3771 				return (1);
3772 		}
3773 		return (0);
3774 	}
3775 
3776 	if (cmd != MD_HALT_DOIT)
3777 		return (1);
3778 
3779 	for (i = 0; i < md_nunits; i++) {
3780 		mnum = MD_MKMIN(setno, i);
3781 		if ((ui = MDI_UNIT(mnum)) == NULL)
3782 			continue;
3783 		if (ui->ui_opsindex != raid_md_ops.md_selfindex)
3784 			continue;
3785 		reset_raid((mr_unit_t *)MD_UNIT(mnum), mnum, 0);
3786 	}
3787 	return (0);
3788 }
3789 
3790 /*
3791  * NAMES:	raid_close_all_devs
3792  * DESCRIPTION: Close all the devices of the unit.
3793  * PARAMETERS:	mr_unit_t *un - pointer to unit structure
3794  * RETURNS:
3795  */
3796 void
3797 raid_close_all_devs(mr_unit_t *un, int init_pw, int md_cflags)
3798 {
3799 	int		i;
3800 	mr_column_t	*device;
3801 
3802 	for (i = 0; i < un->un_totalcolumncnt; i++) {
3803 		device = &un->un_column[i];
3804 		if (device->un_devflags & MD_RAID_DEV_ISOPEN) {
3805 			ASSERT((device->un_dev != (md_dev64_t)0) &&
3806 			    (device->un_dev != NODEV64));
3807 			if ((device->un_devstate & RCS_OKAY) && init_pw)
3808 				(void) init_pw_area(un, device->un_dev,
3809 				    device->un_pwstart, i);
3810 			md_layered_close(device->un_dev, md_cflags);
3811 			device->un_devflags &= ~MD_RAID_DEV_ISOPEN;
3812 		}
3813 	}
3814 }
3815 
3816 /*
3817  * NAMES:	raid_open_all_devs
3818  * DESCRIPTION: Open all the components (columns) of the device unit.
3819  * PARAMETERS:	mr_unit_t *un - pointer to unit structure
3820  * RETURNS:
3821  */
3822 static int
3823 raid_open_all_devs(mr_unit_t *un, int md_oflags)
3824 {
3825 	minor_t		mnum = MD_SID(un);
3826 	int		i;
3827 	int		not_opened = 0;
3828 	int		commit = 0;
3829 	int		col = -1;
3830 	mr_column_t	*device;
3831 	set_t		setno = MD_MIN2SET(MD_SID(un));
3832 	side_t		side = mddb_getsidenum(setno);
3833 	mdkey_t		key;
3834 	mdi_unit_t	*ui = MDI_UNIT(mnum);
3835 
3836 	ui->ui_tstate &= ~MD_INACCESSIBLE;
3837 
3838 	for (i = 0; i < un->un_totalcolumncnt; i++) {
3839 		md_dev64_t tmpdev;
3840 
3841 		device = &un->un_column[i];
3842 
3843 		if (COLUMN_STATE(un, i) & RCS_ERRED) {
3844 			not_opened++;
3845 			continue;
3846 		}
3847 
3848 		if (device->un_devflags & MD_RAID_DEV_ISOPEN)
3849 			continue;
3850 
3851 		tmpdev = device->un_dev;
3852 		/*
3853 		 * Open by device id
3854 		 */
3855 		key = HOTSPARED(un, i) ?
3856 		    device->un_hs_key : device->un_orig_key;
3857 		if ((md_getmajor(tmpdev) != md_major) &&
3858 		    md_devid_found(setno, side, key) == 1) {
3859 			tmpdev = md_resolve_bydevid(mnum, tmpdev, key);
3860 		}
3861 		if (md_layered_open(mnum, &tmpdev, md_oflags)) {
3862 			device->un_dev = tmpdev;
3863 			not_opened++;
3864 			continue;
3865 		}
3866 		device->un_dev = tmpdev;
3867 		device->un_devflags |= MD_RAID_DEV_ISOPEN;
3868 	}
3869 
3870 	/* if open errors and errored devices are 1 then device can run */
3871 	if (not_opened > 1) {
3872 		cmn_err(CE_WARN,
3873 		    "md: %s failed to open. open error on %s\n",
3874 		    md_shortname(MD_SID(un)),
3875 		    md_devname(MD_UN2SET(un), device->un_orig_dev, NULL, 0));
3876 
3877 		ui->ui_tstate |= MD_INACCESSIBLE;
3878 
3879 		SE_NOTIFY(EC_SVM_STATE, ESC_SVM_OPEN_FAIL, SVM_TAG_METADEVICE,
3880 		    MD_UN2SET(un), MD_SID(un));
3881 
3882 		return (not_opened > 1);
3883 	}
3884 
3885 	for (i = 0; i < un->un_totalcolumncnt; i++) {
3886 		device = &un->un_column[i];
3887 		if (device->un_devflags & MD_RAID_DEV_ISOPEN) {
3888 			if (device->un_devstate & RCS_LAST_ERRED) {
3889 			/*
3890 			 * At this point in time there is a possibility
3891 			 * that errors were the result of a controller
3892 			 * failure with more than a single column on it
3893 			 * so clear out last errored columns and let errors
3894 			 * re-occur is necessary.
3895 			 */
3896 				raid_set_state(un, i, RCS_OKAY, 0);
3897 				commit++;
3898 			}
3899 			continue;
3900 		}
3901 		ASSERT(col == -1);
3902 		col = i;
3903 	}
3904 
3905 	if (col != -1) {
3906 		raid_set_state(un, col, RCS_ERRED, 0);
3907 		commit++;
3908 	}
3909 
3910 	if (commit)
3911 		raid_commit(un, NULL);
3912 
3913 	if (col != -1) {
3914 		if (COLUMN_STATE(un, col) & RCS_ERRED) {
3915 			SE_NOTIFY(EC_SVM_STATE, ESC_SVM_ERRED,
3916 			    SVM_TAG_METADEVICE, MD_UN2SET(un), MD_SID(un));
3917 		} else if (COLUMN_STATE(un, col) & RCS_LAST_ERRED) {
3918 			SE_NOTIFY(EC_SVM_STATE, ESC_SVM_LASTERRED,
3919 			    SVM_TAG_METADEVICE, MD_UN2SET(un), MD_SID(un));
3920 		}
3921 	}
3922 
3923 	return (0);
3924 }
3925 
3926 /*
3927  * NAMES:	raid_internal_open
3928  * DESCRIPTION: Do the actual RAID open
3929  * PARAMETERS:	minor_t mnum - minor number of the RAID device
3930  *		int flag -
3931  *		int otyp -
3932  *		int md_oflags - RAID open flags
3933  * RETURNS:	0 if successful, nonzero otherwise
3934  */
3935 int
3936 raid_internal_open(minor_t mnum, int flag, int otyp, int md_oflags)
3937 {
3938 	mr_unit_t	*un;
3939 	mdi_unit_t	*ui;
3940 	int		err = 0;
3941 	int		replay_error = 0;
3942 
3943 	ui = MDI_UNIT(mnum);
3944 	ASSERT(ui != NULL);
3945 
3946 	un = (mr_unit_t *)md_unit_openclose_enter(ui);
3947 	/*
3948 	 * this MUST be checked before md_unit_isopen is checked.
3949 	 * raid_init_columns sets md_unit_isopen to block reset, halt.
3950 	 */
3951 	if ((UNIT_STATE(un) & (RUS_INIT | RUS_DOI)) &&
3952 	    !(md_oflags & MD_OFLG_ISINIT)) {
3953 		md_unit_openclose_exit(ui);
3954 		return (EAGAIN);
3955 	}
3956 
3957 	if ((md_oflags & MD_OFLG_ISINIT) || md_unit_isopen(ui)) {
3958 		err = md_unit_incopen(mnum, flag, otyp);
3959 		goto out;
3960 	}
3961 
3962 	md_unit_readerexit(ui);
3963 
3964 	un = (mr_unit_t *)md_unit_writerlock(ui);
3965 	if (raid_open_all_devs(un, md_oflags) == 0) {
3966 		if ((err = md_unit_incopen(mnum, flag, otyp)) != 0) {
3967 			md_unit_writerexit(ui);
3968 			un = (mr_unit_t *)md_unit_readerlock(ui);
3969 			raid_close_all_devs(un, 0, md_oflags);
3970 			goto out;
3971 		}
3972 	} else {
3973 		/*
3974 		 * if this unit contains more than two errored components
3975 		 * should return error and close all opened devices
3976 		 */
3977 
3978 		md_unit_writerexit(ui);
3979 		un = (mr_unit_t *)md_unit_readerlock(ui);
3980 		raid_close_all_devs(un, 0, md_oflags);
3981 		md_unit_openclose_exit(ui);
3982 		SE_NOTIFY(EC_SVM_STATE, ESC_SVM_OPEN_FAIL, SVM_TAG_METADEVICE,
3983 		    MD_UN2SET(un), MD_SID(un));
3984 		return (ENXIO);
3985 	}
3986 
3987 	if (!(MD_STATUS(un) & MD_UN_REPLAYED)) {
3988 		replay_error = raid_replay(un);
3989 		MD_STATUS(un) |= MD_UN_REPLAYED;
3990 	}
3991 
3992 	md_unit_writerexit(ui);
3993 	un = (mr_unit_t *)md_unit_readerlock(ui);
3994 
3995 	if ((replay_error == RAID_RPLY_READONLY) &&
3996 	    ((flag & (FREAD | FWRITE)) == FREAD)) {
3997 		md_unit_openclose_exit(ui);
3998 		return (0);
3999 	}
4000 
4001 	/* allocate hotspare if possible */
4002 	(void) raid_hotspares();
4003 
4004 
4005 out:
4006 	md_unit_openclose_exit(ui);
4007 	return (err);
4008 }
4009 /*
4010  * NAMES:	raid_open
4011  * DESCRIPTION: RAID metadevice OPEN entry point
4012  * PARAMETERS:	dev_t dev -
4013  *		int flag -
4014  *		int otyp -
4015  *		cred_t * cred_p -
4016  *		int md_oflags -
4017  * RETURNS:
4018  */
4019 /*ARGSUSED1*/
4020 static int
4021 raid_open(dev_t *dev, int flag, int otyp, cred_t *cred_p, int md_oflags)
4022 {
4023 	int		error = 0;
4024 
4025 	if (error = raid_internal_open(getminor(*dev), flag, otyp, md_oflags)) {
4026 		return (error);
4027 	}
4028 	return (0);
4029 }
4030 
4031 /*
4032  * NAMES:	raid_internal_close
4033  * DESCRIPTION: RAID metadevice CLOSE actual implementation
4034  * PARAMETERS:	minor_t - minor number of the RAID device
4035  *		int otyp -
4036  *		int init_pw -
4037  *		int md_cflags - RAID close flags
4038  * RETURNS:	0 if successful, nonzero otherwise
4039  */
4040 /*ARGSUSED*/
4041 int
4042 raid_internal_close(minor_t mnum, int otyp, int init_pw, int md_cflags)
4043 {
4044 	mdi_unit_t	*ui = MDI_UNIT(mnum);
4045 	mr_unit_t	*un;
4046 	int		err = 0;
4047 
4048 	/* single thread */
4049 	un = (mr_unit_t *)md_unit_openclose_enter(ui);
4050 
4051 	/* count closed */
4052 	if ((err = md_unit_decopen(mnum, otyp)) != 0)
4053 		goto out;
4054 	/* close devices, if necessary */
4055 	if (! md_unit_isopen(ui) || (md_cflags & MD_OFLG_PROBEDEV)) {
4056 		raid_close_all_devs(un, init_pw, md_cflags);
4057 	}
4058 
4059 	/* unlock, return success */
4060 out:
4061 	md_unit_openclose_exit(ui);
4062 	return (err);
4063 }
4064 
4065 /*
4066  * NAMES:	raid_close
4067  * DESCRIPTION: RAID metadevice close entry point
4068  * PARAMETERS:	dev_t dev -
4069  *		int flag -
4070  *		int otyp -
4071  *		cred_t * cred_p -
4072  *		int md_oflags -
4073  * RETURNS:
4074  */
4075 /*ARGSUSED1*/
4076 static int
4077 raid_close(dev_t dev, int flag, int otyp, cred_t *cred_p, int md_cflags)
4078 {
4079 	int retval;
4080 
4081 	(void) md_io_writerlock(MDI_UNIT(getminor(dev)));
4082 	retval = raid_internal_close(getminor(dev), otyp, 1, md_cflags);
4083 	(void) md_io_writerexit(MDI_UNIT(getminor(dev)));
4084 	return (retval);
4085 }
4086 
4087 /*
4088  * raid_probe_close_all_devs
4089  */
4090 void
4091 raid_probe_close_all_devs(mr_unit_t *un)
4092 {
4093 	int		i;
4094 	mr_column_t	*device;
4095 
4096 	for (i = 0; i < un->un_totalcolumncnt; i++) {
4097 		device = &un->un_column[i];
4098 
4099 		if (device->un_devflags & MD_RAID_DEV_PROBEOPEN) {
4100 			md_layered_close(device->un_dev,
4101 			    MD_OFLG_PROBEDEV);
4102 			device->un_devflags &= ~MD_RAID_DEV_PROBEOPEN;
4103 		}
4104 	}
4105 }
4106 /*
4107  * Raid_probe_dev:
4108  *
4109  * On entry the unit writerlock is held
4110  */
4111 static int
4112 raid_probe_dev(mdi_unit_t *ui, minor_t mnum)
4113 {
4114 	mr_unit_t	*un;
4115 	int		i;
4116 	int		not_opened = 0;
4117 	int		commit = 0;
4118 	int		col = -1;
4119 	mr_column_t	*device;
4120 	int		md_devopen = 0;
4121 
4122 	if (md_unit_isopen(ui))
4123 		md_devopen++;
4124 
4125 	un = MD_UNIT(mnum);
4126 	/*
4127 	 * If the state has been set to LAST_ERRED because
4128 	 * of an error when the raid device was open at some
4129 	 * point in the past, don't probe. We really don't want
4130 	 * to reset the state in this case.
4131 	 */
4132 	if (UNIT_STATE(un) == RUS_LAST_ERRED)
4133 		return (0);
4134 
4135 	ui->ui_tstate &= ~MD_INACCESSIBLE;
4136 
4137 	for (i = 0; i < un->un_totalcolumncnt; i++) {
4138 		md_dev64_t tmpdev;
4139 
4140 		device = &un->un_column[i];
4141 		if (COLUMN_STATE(un, i) & RCS_ERRED) {
4142 			not_opened++;
4143 			continue;
4144 		}
4145 
4146 		tmpdev = device->un_dev;
4147 		/*
4148 		 * Currently the flags passed are not needed since
4149 		 * there cannot be an underlying metadevice. However
4150 		 * they are kept here for consistency.
4151 		 *
4152 		 * Open by device id
4153 		 */
4154 		tmpdev = md_resolve_bydevid(mnum, tmpdev, HOTSPARED(un, i)?
4155 		    device->un_hs_key : device->un_orig_key);
4156 		if (md_layered_open(mnum, &tmpdev,
4157 		    MD_OFLG_CONT_ERRS | MD_OFLG_PROBEDEV)) {
4158 			device->un_dev = tmpdev;
4159 			not_opened++;
4160 			continue;
4161 		}
4162 		device->un_dev = tmpdev;
4163 
4164 		device->un_devflags |= MD_RAID_DEV_PROBEOPEN;
4165 	}
4166 
4167 	/*
4168 	 * The code below is careful on setting the LAST_ERRED state.
4169 	 *
4170 	 * If open errors and exactly one device has failed we can run.
4171 	 * If more then one device fails we have to figure out when to set
4172 	 * LAST_ERRED state.  The rationale is to avoid unnecessary resyncs
4173 	 * since they are painful and time consuming.
4174 	 *
4175 	 * When more than one component/column fails there are 2 scenerios.
4176 	 *
4177 	 * 1. Metadevice has NOT been opened: In this case, the behavior
4178 	 *    mimics the open symantics. ie. Only the first failed device
4179 	 *    is ERRED and LAST_ERRED is not set.
4180 	 *
4181 	 * 2. Metadevice has been opened: Here the read/write sematics are
4182 	 *    followed. The first failed devicce is ERRED and on the next
4183 	 *    failed device LAST_ERRED is set.
4184 	 */
4185 
4186 	if (not_opened > 1 && !md_devopen) {
4187 		cmn_err(CE_WARN,
4188 		    "md: %s failed to open. open error on %s\n",
4189 		    md_shortname(MD_SID(un)),
4190 		    md_devname(MD_UN2SET(un), device->un_orig_dev, NULL, 0));
4191 		SE_NOTIFY(EC_SVM_STATE, ESC_SVM_OPEN_FAIL, SVM_TAG_METADEVICE,
4192 		    MD_UN2SET(un), MD_SID(un));
4193 		raid_probe_close_all_devs(un);
4194 		ui->ui_tstate |= MD_INACCESSIBLE;
4195 		return (not_opened > 1);
4196 	}
4197 
4198 	if (!md_devopen) {
4199 		for (i = 0; i < un->un_totalcolumncnt; i++) {
4200 			device = &un->un_column[i];
4201 			if (device->un_devflags & MD_RAID_DEV_PROBEOPEN) {
4202 				if (device->un_devstate & RCS_LAST_ERRED) {
4203 					/*
4204 					 * At this point in time there is a
4205 					 * possibility that errors were the
4206 					 * result of a controller failure with
4207 					 * more than a single column on it so
4208 					 * clear out last errored columns and
4209 					 * let errors re-occur is necessary.
4210 					 */
4211 					raid_set_state(un, i, RCS_OKAY, 0);
4212 					commit++;
4213 					}
4214 				continue;
4215 			}
4216 			ASSERT(col == -1);
4217 			/*
4218 			 * note if multiple devices are failing then only
4219 			 * the last one is marked as error
4220 			 */
4221 			col = i;
4222 		}
4223 
4224 		if (col != -1) {
4225 			raid_set_state(un, col, RCS_ERRED, 0);
4226 			commit++;
4227 		}
4228 
4229 	} else {
4230 		for (i = 0; i < un->un_totalcolumncnt; i++) {
4231 			device = &un->un_column[i];
4232 
4233 			/* if we have LAST_ERRED go ahead and commit. */
4234 			if (un->un_state & RUS_LAST_ERRED)
4235 				break;
4236 			/*
4237 			 * could not open the component
4238 			 */
4239 
4240 			if (!(device->un_devflags & MD_RAID_DEV_PROBEOPEN)) {
4241 				col = i;
4242 				raid_set_state(un, col, RCS_ERRED, 0);
4243 				commit++;
4244 			}
4245 		}
4246 	}
4247 
4248 	if (commit)
4249 		raid_commit(un, NULL);
4250 
4251 	if (col != -1) {
4252 		if (COLUMN_STATE(un, col) & RCS_ERRED) {
4253 			SE_NOTIFY(EC_SVM_STATE, ESC_SVM_ERRED,
4254 			    SVM_TAG_METADEVICE, MD_UN2SET(un), MD_SID(un));
4255 		} else if (COLUMN_STATE(un, col) & RCS_LAST_ERRED) {
4256 			SE_NOTIFY(EC_SVM_STATE, ESC_SVM_LASTERRED,
4257 			    SVM_TAG_METADEVICE, MD_UN2SET(un), MD_SID(un));
4258 		}
4259 	}
4260 
4261 	raid_probe_close_all_devs(un);
4262 	return (0);
4263 }
4264 
4265 static int
4266 raid_imp_set(
4267 	set_t	setno
4268 )
4269 {
4270 	mddb_recid_t    recid;
4271 	int		i, gotsomething;
4272 	mddb_type_t	typ1;
4273 	mddb_de_ic_t	*dep;
4274 	mddb_rb32_t	*rbp;
4275 	mr_unit_t	*un64;
4276 	mr_unit32_od_t	*un32;
4277 	md_dev64_t	self_devt;
4278 	minor_t		*self_id;	/* minor needs to be updated */
4279 	md_parent_t	*parent_id;	/* parent needs to be updated */
4280 	mddb_recid_t	*record_id;	 /* record id needs to be updated */
4281 	hsp_t		*hsp_id;
4282 
4283 	gotsomething = 0;
4284 
4285 	typ1 = (mddb_type_t)md_getshared_key(setno,
4286 	    raid_md_ops.md_driver.md_drivername);
4287 	recid = mddb_makerecid(setno, 0);
4288 
4289 	while ((recid = mddb_getnextrec(recid, typ1, 0)) > 0) {
4290 		if (mddb_getrecprivate(recid) & MD_PRV_GOTIT)
4291 			continue;
4292 
4293 		dep = mddb_getrecdep(recid);
4294 		rbp = dep->de_rb;
4295 
4296 		switch (rbp->rb_revision) {
4297 		case MDDB_REV_RB:
4298 		case MDDB_REV_RBFN:
4299 			/*
4300 			 * Small device
4301 			 */
4302 			un32 = (mr_unit32_od_t *)mddb_getrecaddr(recid);
4303 			self_id = &(un32->c.un_self_id);
4304 			parent_id = &(un32->c.un_parent);
4305 			record_id = &(un32->c.un_record_id);
4306 			hsp_id = &(un32->un_hsp_id);
4307 
4308 			for (i = 0; i < un32->un_totalcolumncnt; i++) {
4309 				mr_column32_od_t *device;
4310 
4311 				device = &un32->un_column[i];
4312 				if (!md_update_minor(setno, mddb_getsidenum
4313 				    (setno), device->un_orig_key))
4314 					goto out;
4315 
4316 				if (device->un_hs_id != 0)
4317 					device->un_hs_id =
4318 					    MAKERECID(setno, device->un_hs_id);
4319 			}
4320 			break;
4321 		case MDDB_REV_RB64:
4322 		case MDDB_REV_RB64FN:
4323 			un64 = (mr_unit_t *)mddb_getrecaddr(recid);
4324 			self_id = &(un64->c.un_self_id);
4325 			parent_id = &(un64->c.un_parent);
4326 			record_id = &(un64->c.un_record_id);
4327 			hsp_id = &(un64->un_hsp_id);
4328 
4329 			for (i = 0; i < un64->un_totalcolumncnt; i++) {
4330 				mr_column_t	*device;
4331 
4332 				device = &un64->un_column[i];
4333 				if (!md_update_minor(setno, mddb_getsidenum
4334 				    (setno), device->un_orig_key))
4335 					goto out;
4336 
4337 				if (device->un_hs_id != 0)
4338 					device->un_hs_id =
4339 					    MAKERECID(setno, device->un_hs_id);
4340 			}
4341 			break;
4342 		}
4343 
4344 		/*
4345 		 * If this is a top level and a friendly name metadevice,
4346 		 * update its minor in the namespace.
4347 		 */
4348 		if ((*parent_id == MD_NO_PARENT) &&
4349 		    ((rbp->rb_revision == MDDB_REV_RBFN) ||
4350 		    (rbp->rb_revision == MDDB_REV_RB64FN))) {
4351 
4352 			self_devt = md_makedevice(md_major, *self_id);
4353 			if (!md_update_top_device_minor(setno,
4354 			    mddb_getsidenum(setno), self_devt))
4355 				goto out;
4356 		}
4357 
4358 		/*
4359 		 * Update unit with the imported setno
4360 		 */
4361 		mddb_setrecprivate(recid, MD_PRV_GOTIT);
4362 
4363 		*self_id = MD_MKMIN(setno, MD_MIN2UNIT(*self_id));
4364 
4365 		if (*hsp_id != -1)
4366 			*hsp_id = MAKERECID(setno, DBID(*hsp_id));
4367 
4368 		if (*parent_id != MD_NO_PARENT)
4369 			*parent_id = MD_MKMIN(setno, MD_MIN2UNIT(*parent_id));
4370 		*record_id = MAKERECID(setno, DBID(*record_id));
4371 		gotsomething = 1;
4372 	}
4373 
4374 out:
4375 	return (gotsomething);
4376 }
4377 
4378 static md_named_services_t raid_named_services[] = {
4379 	{raid_hotspares,			"poke hotspares"	},
4380 	{raid_rename_check,			MDRNM_CHECK		},
4381 	{raid_rename_lock,			MDRNM_LOCK		},
4382 	{(intptr_t (*)()) raid_rename_unlock,	MDRNM_UNLOCK		},
4383 	{(intptr_t (*)()) raid_probe_dev,	"probe open test"	},
4384 	{NULL,					0			}
4385 };
4386 
4387 md_ops_t raid_md_ops = {
4388 	raid_open,		/* open */
4389 	raid_close,		/* close */
4390 	md_raid_strategy,	/* strategy */
4391 	NULL,			/* print */
4392 	NULL,			/* dump */
4393 	NULL,			/* read */
4394 	NULL,			/* write */
4395 	md_raid_ioctl,		/* ioctl, */
4396 	raid_snarf,		/* raid_snarf */
4397 	raid_halt,		/* raid_halt */
4398 	NULL,			/* aread */
4399 	NULL,			/* awrite */
4400 	raid_imp_set,		/* import set */
4401 	raid_named_services
4402 };
4403 
4404 static void
4405 init_init()
4406 {
4407 	/* default to a second */
4408 	if (md_wr_wait == 0)
4409 		md_wr_wait = md_hz >> 1;
4410 
4411 	raid_parent_cache = kmem_cache_create("md_raid_parent",
4412 	    sizeof (md_raidps_t), 0, raid_parent_constructor,
4413 	    raid_parent_destructor, raid_run_queue, NULL, NULL, 0);
4414 	raid_child_cache = kmem_cache_create("md_raid_child",
4415 	    sizeof (md_raidcs_t) - sizeof (buf_t) + biosize(), 0,
4416 	    raid_child_constructor, raid_child_destructor,
4417 	    raid_run_queue, NULL, NULL, 0);
4418 	raid_cbuf_cache = kmem_cache_create("md_raid_cbufs",
4419 	    sizeof (md_raidcbuf_t), 0, raid_cbuf_constructor,
4420 	    raid_cbuf_destructor, raid_run_queue, NULL, NULL, 0);
4421 }
4422 
4423 static void
4424 fini_uninit()
4425 {
4426 	kmem_cache_destroy(raid_parent_cache);
4427 	kmem_cache_destroy(raid_child_cache);
4428 	kmem_cache_destroy(raid_cbuf_cache);
4429 	raid_parent_cache = raid_child_cache = raid_cbuf_cache = NULL;
4430 }
4431 
4432 /* define the module linkage */
4433 MD_PLUGIN_MISC_MODULE("raid module", init_init(), fini_uninit())
4434